150
|
1 //===--- CodeGenFunction.cpp - Emit LLVM Code from ASTs for a Function ----===//
|
|
2 //
|
|
3 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
|
|
4 // See https://llvm.org/LICENSE.txt for license information.
|
|
5 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
|
|
6 //
|
|
7 //===----------------------------------------------------------------------===//
|
|
8 //
|
|
9 // This coordinates the per-function state used while generating code.
|
|
10 //
|
|
11 //===----------------------------------------------------------------------===//
|
|
12
|
|
13 #include "CodeGenFunction.h"
|
|
14 #include "CGBlocks.h"
|
|
15 #include "CGCUDARuntime.h"
|
|
16 #include "CGCXXABI.h"
|
|
17 #include "CGCleanup.h"
|
|
18 #include "CGDebugInfo.h"
|
|
19 #include "CGOpenMPRuntime.h"
|
|
20 #include "CodeGenModule.h"
|
|
21 #include "CodeGenPGO.h"
|
|
22 #include "TargetInfo.h"
|
|
23 #include "clang/AST/ASTContext.h"
|
|
24 #include "clang/AST/ASTLambda.h"
|
|
25 #include "clang/AST/Attr.h"
|
|
26 #include "clang/AST/Decl.h"
|
|
27 #include "clang/AST/DeclCXX.h"
|
|
28 #include "clang/AST/StmtCXX.h"
|
|
29 #include "clang/AST/StmtObjC.h"
|
|
30 #include "clang/Basic/Builtins.h"
|
|
31 #include "clang/Basic/CodeGenOptions.h"
|
|
32 #include "clang/Basic/TargetInfo.h"
|
|
33 #include "clang/CodeGen/CGFunctionInfo.h"
|
|
34 #include "clang/Frontend/FrontendDiagnostic.h"
|
|
35 #include "llvm/Frontend/OpenMP/OMPIRBuilder.h"
|
|
36 #include "llvm/IR/DataLayout.h"
|
|
37 #include "llvm/IR/Dominators.h"
|
|
38 #include "llvm/IR/FPEnv.h"
|
|
39 #include "llvm/IR/IntrinsicInst.h"
|
|
40 #include "llvm/IR/Intrinsics.h"
|
|
41 #include "llvm/IR/MDBuilder.h"
|
|
42 #include "llvm/IR/Operator.h"
|
|
43 #include "llvm/Transforms/Utils/PromoteMemToReg.h"
|
|
44 using namespace clang;
|
|
45 using namespace CodeGen;
|
|
46
|
|
47 /// shouldEmitLifetimeMarkers - Decide whether we need emit the life-time
|
|
48 /// markers.
|
|
49 static bool shouldEmitLifetimeMarkers(const CodeGenOptions &CGOpts,
|
|
50 const LangOptions &LangOpts) {
|
|
51 if (CGOpts.DisableLifetimeMarkers)
|
|
52 return false;
|
|
53
|
|
54 // Sanitizers may use markers.
|
|
55 if (CGOpts.SanitizeAddressUseAfterScope ||
|
|
56 LangOpts.Sanitize.has(SanitizerKind::HWAddress) ||
|
|
57 LangOpts.Sanitize.has(SanitizerKind::Memory))
|
|
58 return true;
|
|
59
|
|
60 // For now, only in optimized builds.
|
|
61 return CGOpts.OptimizationLevel != 0;
|
|
62 }
|
|
63
|
|
64 CodeGenFunction::CodeGenFunction(CodeGenModule &cgm, bool suppressNewContext)
|
|
65 : CodeGenTypeCache(cgm), CGM(cgm), Target(cgm.getTarget()),
|
|
66 Builder(cgm, cgm.getModule().getContext(), llvm::ConstantFolder(),
|
|
67 CGBuilderInserterTy(this)),
|
|
68 SanOpts(CGM.getLangOpts().Sanitize), DebugInfo(CGM.getModuleDebugInfo()),
|
|
69 PGO(cgm), ShouldEmitLifetimeMarkers(shouldEmitLifetimeMarkers(
|
|
70 CGM.getCodeGenOpts(), CGM.getLangOpts())) {
|
|
71 if (!suppressNewContext)
|
|
72 CGM.getCXXABI().getMangleContext().startNewFunction();
|
|
73
|
|
74 llvm::FastMathFlags FMF;
|
|
75 if (CGM.getLangOpts().FastMath)
|
|
76 FMF.setFast();
|
|
77 if (CGM.getLangOpts().FiniteMathOnly) {
|
|
78 FMF.setNoNaNs();
|
|
79 FMF.setNoInfs();
|
|
80 }
|
|
81 if (CGM.getCodeGenOpts().NoNaNsFPMath) {
|
|
82 FMF.setNoNaNs();
|
|
83 }
|
|
84 if (CGM.getCodeGenOpts().NoSignedZeros) {
|
|
85 FMF.setNoSignedZeros();
|
|
86 }
|
|
87 if (CGM.getCodeGenOpts().ReciprocalMath) {
|
|
88 FMF.setAllowReciprocal();
|
|
89 }
|
|
90 if (CGM.getCodeGenOpts().Reassociate) {
|
|
91 FMF.setAllowReassoc();
|
|
92 }
|
|
93 Builder.setFastMathFlags(FMF);
|
|
94 SetFPModel();
|
|
95 }
|
|
96
|
|
97 CodeGenFunction::~CodeGenFunction() {
|
|
98 assert(LifetimeExtendedCleanupStack.empty() && "failed to emit a cleanup");
|
|
99
|
|
100 // If there are any unclaimed block infos, go ahead and destroy them
|
|
101 // now. This can happen if IR-gen gets clever and skips evaluating
|
|
102 // something.
|
|
103 if (FirstBlockInfo)
|
|
104 destroyBlockInfos(FirstBlockInfo);
|
|
105
|
|
106 if (getLangOpts().OpenMP && CurFn)
|
|
107 CGM.getOpenMPRuntime().functionFinished(*this);
|
|
108
|
|
109 // If we have an OpenMPIRBuilder we want to finalize functions (incl.
|
|
110 // outlining etc) at some point. Doing it once the function codegen is done
|
|
111 // seems to be a reasonable spot. We do it here, as opposed to the deletion
|
|
112 // time of the CodeGenModule, because we have to ensure the IR has not yet
|
|
113 // been "emitted" to the outside, thus, modifications are still sensible.
|
|
114 if (llvm::OpenMPIRBuilder *OMPBuilder = CGM.getOpenMPIRBuilder())
|
|
115 OMPBuilder->finalize();
|
|
116 }
|
|
117
|
|
118 // Map the LangOption for exception behavior into
|
|
119 // the corresponding enum in the IR.
|
173
|
120 llvm::fp::ExceptionBehavior
|
|
121 clang::ToConstrainedExceptMD(LangOptions::FPExceptionModeKind Kind) {
|
150
|
122
|
|
123 switch (Kind) {
|
|
124 case LangOptions::FPE_Ignore: return llvm::fp::ebIgnore;
|
|
125 case LangOptions::FPE_MayTrap: return llvm::fp::ebMayTrap;
|
|
126 case LangOptions::FPE_Strict: return llvm::fp::ebStrict;
|
|
127 }
|
|
128 llvm_unreachable("Unsupported FP Exception Behavior");
|
|
129 }
|
|
130
|
|
131 void CodeGenFunction::SetFPModel() {
|
173
|
132 llvm::RoundingMode RM = getLangOpts().getFPRoundingMode();
|
150
|
133 auto fpExceptionBehavior = ToConstrainedExceptMD(
|
|
134 getLangOpts().getFPExceptionMode());
|
|
135
|
173
|
136 Builder.setDefaultConstrainedRounding(RM);
|
|
137 Builder.setDefaultConstrainedExcept(fpExceptionBehavior);
|
|
138 Builder.setIsFPConstrained(fpExceptionBehavior != llvm::fp::ebIgnore ||
|
|
139 RM != llvm::RoundingMode::NearestTiesToEven);
|
150
|
140 }
|
|
141
|
|
142 LValue CodeGenFunction::MakeNaturalAlignAddrLValue(llvm::Value *V, QualType T) {
|
|
143 LValueBaseInfo BaseInfo;
|
|
144 TBAAAccessInfo TBAAInfo;
|
173
|
145 CharUnits Alignment = CGM.getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo);
|
150
|
146 return LValue::MakeAddr(Address(V, Alignment), T, getContext(), BaseInfo,
|
|
147 TBAAInfo);
|
|
148 }
|
|
149
|
|
150 /// Given a value of type T* that may not be to a complete object,
|
|
151 /// construct an l-value with the natural pointee alignment of T.
|
|
152 LValue
|
|
153 CodeGenFunction::MakeNaturalAlignPointeeAddrLValue(llvm::Value *V, QualType T) {
|
|
154 LValueBaseInfo BaseInfo;
|
|
155 TBAAAccessInfo TBAAInfo;
|
173
|
156 CharUnits Align = CGM.getNaturalTypeAlignment(T, &BaseInfo, &TBAAInfo,
|
|
157 /* forPointeeType= */ true);
|
150
|
158 return MakeAddrLValue(Address(V, Align), T, BaseInfo, TBAAInfo);
|
|
159 }
|
|
160
|
|
161
|
|
162 llvm::Type *CodeGenFunction::ConvertTypeForMem(QualType T) {
|
|
163 return CGM.getTypes().ConvertTypeForMem(T);
|
|
164 }
|
|
165
|
|
166 llvm::Type *CodeGenFunction::ConvertType(QualType T) {
|
|
167 return CGM.getTypes().ConvertType(T);
|
|
168 }
|
|
169
|
|
170 TypeEvaluationKind CodeGenFunction::getEvaluationKind(QualType type) {
|
|
171 type = type.getCanonicalType();
|
|
172 while (true) {
|
|
173 switch (type->getTypeClass()) {
|
|
174 #define TYPE(name, parent)
|
|
175 #define ABSTRACT_TYPE(name, parent)
|
|
176 #define NON_CANONICAL_TYPE(name, parent) case Type::name:
|
|
177 #define DEPENDENT_TYPE(name, parent) case Type::name:
|
|
178 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(name, parent) case Type::name:
|
|
179 #include "clang/AST/TypeNodes.inc"
|
|
180 llvm_unreachable("non-canonical or dependent type in IR-generation");
|
|
181
|
|
182 case Type::Auto:
|
|
183 case Type::DeducedTemplateSpecialization:
|
|
184 llvm_unreachable("undeduced type in IR-generation");
|
|
185
|
|
186 // Various scalar types.
|
|
187 case Type::Builtin:
|
|
188 case Type::Pointer:
|
|
189 case Type::BlockPointer:
|
|
190 case Type::LValueReference:
|
|
191 case Type::RValueReference:
|
|
192 case Type::MemberPointer:
|
|
193 case Type::Vector:
|
|
194 case Type::ExtVector:
|
173
|
195 case Type::ConstantMatrix:
|
150
|
196 case Type::FunctionProto:
|
|
197 case Type::FunctionNoProto:
|
|
198 case Type::Enum:
|
|
199 case Type::ObjCObjectPointer:
|
|
200 case Type::Pipe:
|
173
|
201 case Type::ExtInt:
|
150
|
202 return TEK_Scalar;
|
|
203
|
|
204 // Complexes.
|
|
205 case Type::Complex:
|
|
206 return TEK_Complex;
|
|
207
|
|
208 // Arrays, records, and Objective-C objects.
|
|
209 case Type::ConstantArray:
|
|
210 case Type::IncompleteArray:
|
|
211 case Type::VariableArray:
|
|
212 case Type::Record:
|
|
213 case Type::ObjCObject:
|
|
214 case Type::ObjCInterface:
|
|
215 return TEK_Aggregate;
|
|
216
|
|
217 // We operate on atomic values according to their underlying type.
|
|
218 case Type::Atomic:
|
|
219 type = cast<AtomicType>(type)->getValueType();
|
|
220 continue;
|
|
221 }
|
|
222 llvm_unreachable("unknown type kind!");
|
|
223 }
|
|
224 }
|
|
225
|
|
226 llvm::DebugLoc CodeGenFunction::EmitReturnBlock() {
|
|
227 // For cleanliness, we try to avoid emitting the return block for
|
|
228 // simple cases.
|
|
229 llvm::BasicBlock *CurBB = Builder.GetInsertBlock();
|
|
230
|
|
231 if (CurBB) {
|
|
232 assert(!CurBB->getTerminator() && "Unexpected terminated block.");
|
|
233
|
|
234 // We have a valid insert point, reuse it if it is empty or there are no
|
|
235 // explicit jumps to the return block.
|
|
236 if (CurBB->empty() || ReturnBlock.getBlock()->use_empty()) {
|
|
237 ReturnBlock.getBlock()->replaceAllUsesWith(CurBB);
|
|
238 delete ReturnBlock.getBlock();
|
|
239 ReturnBlock = JumpDest();
|
|
240 } else
|
|
241 EmitBlock(ReturnBlock.getBlock());
|
|
242 return llvm::DebugLoc();
|
|
243 }
|
|
244
|
|
245 // Otherwise, if the return block is the target of a single direct
|
|
246 // branch then we can just put the code in that block instead. This
|
|
247 // cleans up functions which started with a unified return block.
|
|
248 if (ReturnBlock.getBlock()->hasOneUse()) {
|
|
249 llvm::BranchInst *BI =
|
|
250 dyn_cast<llvm::BranchInst>(*ReturnBlock.getBlock()->user_begin());
|
|
251 if (BI && BI->isUnconditional() &&
|
|
252 BI->getSuccessor(0) == ReturnBlock.getBlock()) {
|
|
253 // Record/return the DebugLoc of the simple 'return' expression to be used
|
|
254 // later by the actual 'ret' instruction.
|
|
255 llvm::DebugLoc Loc = BI->getDebugLoc();
|
|
256 Builder.SetInsertPoint(BI->getParent());
|
|
257 BI->eraseFromParent();
|
|
258 delete ReturnBlock.getBlock();
|
|
259 ReturnBlock = JumpDest();
|
|
260 return Loc;
|
|
261 }
|
|
262 }
|
|
263
|
|
264 // FIXME: We are at an unreachable point, there is no reason to emit the block
|
|
265 // unless it has uses. However, we still need a place to put the debug
|
|
266 // region.end for now.
|
|
267
|
|
268 EmitBlock(ReturnBlock.getBlock());
|
|
269 return llvm::DebugLoc();
|
|
270 }
|
|
271
|
|
272 static void EmitIfUsed(CodeGenFunction &CGF, llvm::BasicBlock *BB) {
|
|
273 if (!BB) return;
|
|
274 if (!BB->use_empty())
|
|
275 return CGF.CurFn->getBasicBlockList().push_back(BB);
|
|
276 delete BB;
|
|
277 }
|
|
278
|
|
279 void CodeGenFunction::FinishFunction(SourceLocation EndLoc) {
|
|
280 assert(BreakContinueStack.empty() &&
|
|
281 "mismatched push/pop in break/continue stack!");
|
|
282
|
|
283 bool OnlySimpleReturnStmts = NumSimpleReturnExprs > 0
|
|
284 && NumSimpleReturnExprs == NumReturnExprs
|
|
285 && ReturnBlock.getBlock()->use_empty();
|
|
286 // Usually the return expression is evaluated before the cleanup
|
|
287 // code. If the function contains only a simple return statement,
|
|
288 // such as a constant, the location before the cleanup code becomes
|
|
289 // the last useful breakpoint in the function, because the simple
|
|
290 // return expression will be evaluated after the cleanup code. To be
|
|
291 // safe, set the debug location for cleanup code to the location of
|
|
292 // the return statement. Otherwise the cleanup code should be at the
|
|
293 // end of the function's lexical scope.
|
|
294 //
|
|
295 // If there are multiple branches to the return block, the branch
|
|
296 // instructions will get the location of the return statements and
|
|
297 // all will be fine.
|
|
298 if (CGDebugInfo *DI = getDebugInfo()) {
|
|
299 if (OnlySimpleReturnStmts)
|
|
300 DI->EmitLocation(Builder, LastStopPoint);
|
|
301 else
|
|
302 DI->EmitLocation(Builder, EndLoc);
|
|
303 }
|
|
304
|
|
305 // Pop any cleanups that might have been associated with the
|
|
306 // parameters. Do this in whatever block we're currently in; it's
|
|
307 // important to do this before we enter the return block or return
|
|
308 // edges will be *really* confused.
|
|
309 bool HasCleanups = EHStack.stable_begin() != PrologueCleanupDepth;
|
|
310 bool HasOnlyLifetimeMarkers =
|
|
311 HasCleanups && EHStack.containsOnlyLifetimeMarkers(PrologueCleanupDepth);
|
|
312 bool EmitRetDbgLoc = !HasCleanups || HasOnlyLifetimeMarkers;
|
|
313 if (HasCleanups) {
|
|
314 // Make sure the line table doesn't jump back into the body for
|
|
315 // the ret after it's been at EndLoc.
|
|
316 Optional<ApplyDebugLocation> AL;
|
|
317 if (CGDebugInfo *DI = getDebugInfo()) {
|
|
318 if (OnlySimpleReturnStmts)
|
|
319 DI->EmitLocation(Builder, EndLoc);
|
|
320 else
|
|
321 // We may not have a valid end location. Try to apply it anyway, and
|
|
322 // fall back to an artificial location if needed.
|
|
323 AL = ApplyDebugLocation::CreateDefaultArtificial(*this, EndLoc);
|
|
324 }
|
|
325
|
|
326 PopCleanupBlocks(PrologueCleanupDepth);
|
|
327 }
|
|
328
|
|
329 // Emit function epilog (to return).
|
|
330 llvm::DebugLoc Loc = EmitReturnBlock();
|
|
331
|
|
332 if (ShouldInstrumentFunction()) {
|
|
333 if (CGM.getCodeGenOpts().InstrumentFunctions)
|
|
334 CurFn->addFnAttr("instrument-function-exit", "__cyg_profile_func_exit");
|
|
335 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
|
|
336 CurFn->addFnAttr("instrument-function-exit-inlined",
|
|
337 "__cyg_profile_func_exit");
|
|
338 }
|
|
339
|
|
340 // Emit debug descriptor for function end.
|
|
341 if (CGDebugInfo *DI = getDebugInfo())
|
|
342 DI->EmitFunctionEnd(Builder, CurFn);
|
|
343
|
|
344 // Reset the debug location to that of the simple 'return' expression, if any
|
|
345 // rather than that of the end of the function's scope '}'.
|
|
346 ApplyDebugLocation AL(*this, Loc);
|
|
347 EmitFunctionEpilog(*CurFnInfo, EmitRetDbgLoc, EndLoc);
|
|
348 EmitEndEHSpec(CurCodeDecl);
|
|
349
|
|
350 assert(EHStack.empty() &&
|
|
351 "did not remove all scopes from cleanup stack!");
|
|
352
|
|
353 // If someone did an indirect goto, emit the indirect goto block at the end of
|
|
354 // the function.
|
|
355 if (IndirectBranch) {
|
|
356 EmitBlock(IndirectBranch->getParent());
|
|
357 Builder.ClearInsertionPoint();
|
|
358 }
|
|
359
|
|
360 // If some of our locals escaped, insert a call to llvm.localescape in the
|
|
361 // entry block.
|
|
362 if (!EscapedLocals.empty()) {
|
|
363 // Invert the map from local to index into a simple vector. There should be
|
|
364 // no holes.
|
|
365 SmallVector<llvm::Value *, 4> EscapeArgs;
|
|
366 EscapeArgs.resize(EscapedLocals.size());
|
|
367 for (auto &Pair : EscapedLocals)
|
|
368 EscapeArgs[Pair.second] = Pair.first;
|
|
369 llvm::Function *FrameEscapeFn = llvm::Intrinsic::getDeclaration(
|
|
370 &CGM.getModule(), llvm::Intrinsic::localescape);
|
|
371 CGBuilderTy(*this, AllocaInsertPt).CreateCall(FrameEscapeFn, EscapeArgs);
|
|
372 }
|
|
373
|
|
374 // Remove the AllocaInsertPt instruction, which is just a convenience for us.
|
|
375 llvm::Instruction *Ptr = AllocaInsertPt;
|
|
376 AllocaInsertPt = nullptr;
|
|
377 Ptr->eraseFromParent();
|
|
378
|
|
379 // If someone took the address of a label but never did an indirect goto, we
|
|
380 // made a zero entry PHI node, which is illegal, zap it now.
|
|
381 if (IndirectBranch) {
|
|
382 llvm::PHINode *PN = cast<llvm::PHINode>(IndirectBranch->getAddress());
|
|
383 if (PN->getNumIncomingValues() == 0) {
|
|
384 PN->replaceAllUsesWith(llvm::UndefValue::get(PN->getType()));
|
|
385 PN->eraseFromParent();
|
|
386 }
|
|
387 }
|
|
388
|
|
389 EmitIfUsed(*this, EHResumeBlock);
|
|
390 EmitIfUsed(*this, TerminateLandingPad);
|
|
391 EmitIfUsed(*this, TerminateHandler);
|
|
392 EmitIfUsed(*this, UnreachableBlock);
|
|
393
|
|
394 for (const auto &FuncletAndParent : TerminateFunclets)
|
|
395 EmitIfUsed(*this, FuncletAndParent.second);
|
|
396
|
|
397 if (CGM.getCodeGenOpts().EmitDeclMetadata)
|
|
398 EmitDeclMetadata();
|
|
399
|
|
400 for (SmallVectorImpl<std::pair<llvm::Instruction *, llvm::Value *> >::iterator
|
|
401 I = DeferredReplacements.begin(),
|
|
402 E = DeferredReplacements.end();
|
|
403 I != E; ++I) {
|
|
404 I->first->replaceAllUsesWith(I->second);
|
|
405 I->first->eraseFromParent();
|
|
406 }
|
|
407
|
|
408 // Eliminate CleanupDestSlot alloca by replacing it with SSA values and
|
|
409 // PHIs if the current function is a coroutine. We don't do it for all
|
|
410 // functions as it may result in slight increase in numbers of instructions
|
|
411 // if compiled with no optimizations. We do it for coroutine as the lifetime
|
|
412 // of CleanupDestSlot alloca make correct coroutine frame building very
|
|
413 // difficult.
|
|
414 if (NormalCleanupDest.isValid() && isCoroutine()) {
|
|
415 llvm::DominatorTree DT(*CurFn);
|
|
416 llvm::PromoteMemToReg(
|
|
417 cast<llvm::AllocaInst>(NormalCleanupDest.getPointer()), DT);
|
|
418 NormalCleanupDest = Address::invalid();
|
|
419 }
|
|
420
|
|
421 // Scan function arguments for vector width.
|
|
422 for (llvm::Argument &A : CurFn->args())
|
|
423 if (auto *VT = dyn_cast<llvm::VectorType>(A.getType()))
|
173
|
424 LargestVectorWidth =
|
|
425 std::max((uint64_t)LargestVectorWidth,
|
|
426 VT->getPrimitiveSizeInBits().getKnownMinSize());
|
150
|
427
|
|
428 // Update vector width based on return type.
|
|
429 if (auto *VT = dyn_cast<llvm::VectorType>(CurFn->getReturnType()))
|
173
|
430 LargestVectorWidth =
|
|
431 std::max((uint64_t)LargestVectorWidth,
|
|
432 VT->getPrimitiveSizeInBits().getKnownMinSize());
|
150
|
433
|
|
434 // Add the required-vector-width attribute. This contains the max width from:
|
|
435 // 1. min-vector-width attribute used in the source program.
|
|
436 // 2. Any builtins used that have a vector width specified.
|
|
437 // 3. Values passed in and out of inline assembly.
|
|
438 // 4. Width of vector arguments and return types for this function.
|
|
439 // 5. Width of vector aguments and return types for functions called by this
|
|
440 // function.
|
|
441 CurFn->addFnAttr("min-legal-vector-width", llvm::utostr(LargestVectorWidth));
|
|
442
|
|
443 // If we generated an unreachable return block, delete it now.
|
|
444 if (ReturnBlock.isValid() && ReturnBlock.getBlock()->use_empty()) {
|
|
445 Builder.ClearInsertionPoint();
|
|
446 ReturnBlock.getBlock()->eraseFromParent();
|
|
447 }
|
|
448 if (ReturnValue.isValid()) {
|
|
449 auto *RetAlloca = dyn_cast<llvm::AllocaInst>(ReturnValue.getPointer());
|
|
450 if (RetAlloca && RetAlloca->use_empty()) {
|
|
451 RetAlloca->eraseFromParent();
|
|
452 ReturnValue = Address::invalid();
|
|
453 }
|
|
454 }
|
|
455 }
|
|
456
|
|
457 /// ShouldInstrumentFunction - Return true if the current function should be
|
|
458 /// instrumented with __cyg_profile_func_* calls
|
|
459 bool CodeGenFunction::ShouldInstrumentFunction() {
|
|
460 if (!CGM.getCodeGenOpts().InstrumentFunctions &&
|
|
461 !CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining &&
|
|
462 !CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
|
|
463 return false;
|
|
464 if (!CurFuncDecl || CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>())
|
|
465 return false;
|
|
466 return true;
|
|
467 }
|
|
468
|
|
469 /// ShouldXRayInstrument - Return true if the current function should be
|
|
470 /// instrumented with XRay nop sleds.
|
|
471 bool CodeGenFunction::ShouldXRayInstrumentFunction() const {
|
|
472 return CGM.getCodeGenOpts().XRayInstrumentFunctions;
|
|
473 }
|
|
474
|
|
475 /// AlwaysEmitXRayCustomEvents - Return true if we should emit IR for calls to
|
|
476 /// the __xray_customevent(...) builtin calls, when doing XRay instrumentation.
|
|
477 bool CodeGenFunction::AlwaysEmitXRayCustomEvents() const {
|
|
478 return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
|
|
479 (CGM.getCodeGenOpts().XRayAlwaysEmitCustomEvents ||
|
|
480 CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
|
|
481 XRayInstrKind::Custom);
|
|
482 }
|
|
483
|
|
484 bool CodeGenFunction::AlwaysEmitXRayTypedEvents() const {
|
|
485 return CGM.getCodeGenOpts().XRayInstrumentFunctions &&
|
|
486 (CGM.getCodeGenOpts().XRayAlwaysEmitTypedEvents ||
|
|
487 CGM.getCodeGenOpts().XRayInstrumentationBundle.Mask ==
|
|
488 XRayInstrKind::Typed);
|
|
489 }
|
|
490
|
|
491 llvm::Constant *
|
|
492 CodeGenFunction::EncodeAddrForUseInPrologue(llvm::Function *F,
|
|
493 llvm::Constant *Addr) {
|
|
494 // Addresses stored in prologue data can't require run-time fixups and must
|
|
495 // be PC-relative. Run-time fixups are undesirable because they necessitate
|
|
496 // writable text segments, which are unsafe. And absolute addresses are
|
|
497 // undesirable because they break PIE mode.
|
|
498
|
|
499 // Add a layer of indirection through a private global. Taking its address
|
|
500 // won't result in a run-time fixup, even if Addr has linkonce_odr linkage.
|
|
501 auto *GV = new llvm::GlobalVariable(CGM.getModule(), Addr->getType(),
|
|
502 /*isConstant=*/true,
|
|
503 llvm::GlobalValue::PrivateLinkage, Addr);
|
|
504
|
|
505 // Create a PC-relative address.
|
|
506 auto *GOTAsInt = llvm::ConstantExpr::getPtrToInt(GV, IntPtrTy);
|
|
507 auto *FuncAsInt = llvm::ConstantExpr::getPtrToInt(F, IntPtrTy);
|
|
508 auto *PCRelAsInt = llvm::ConstantExpr::getSub(GOTAsInt, FuncAsInt);
|
|
509 return (IntPtrTy == Int32Ty)
|
|
510 ? PCRelAsInt
|
|
511 : llvm::ConstantExpr::getTrunc(PCRelAsInt, Int32Ty);
|
|
512 }
|
|
513
|
|
514 llvm::Value *
|
|
515 CodeGenFunction::DecodeAddrUsedInPrologue(llvm::Value *F,
|
|
516 llvm::Value *EncodedAddr) {
|
|
517 // Reconstruct the address of the global.
|
|
518 auto *PCRelAsInt = Builder.CreateSExt(EncodedAddr, IntPtrTy);
|
|
519 auto *FuncAsInt = Builder.CreatePtrToInt(F, IntPtrTy, "func_addr.int");
|
|
520 auto *GOTAsInt = Builder.CreateAdd(PCRelAsInt, FuncAsInt, "global_addr.int");
|
|
521 auto *GOTAddr = Builder.CreateIntToPtr(GOTAsInt, Int8PtrPtrTy, "global_addr");
|
|
522
|
|
523 // Load the original pointer through the global.
|
|
524 return Builder.CreateLoad(Address(GOTAddr, getPointerAlign()),
|
|
525 "decoded_addr");
|
|
526 }
|
|
527
|
|
528 void CodeGenFunction::EmitOpenCLKernelMetadata(const FunctionDecl *FD,
|
|
529 llvm::Function *Fn)
|
|
530 {
|
|
531 if (!FD->hasAttr<OpenCLKernelAttr>())
|
|
532 return;
|
|
533
|
|
534 llvm::LLVMContext &Context = getLLVMContext();
|
|
535
|
|
536 CGM.GenOpenCLArgMetadata(Fn, FD, this);
|
|
537
|
|
538 if (const VecTypeHintAttr *A = FD->getAttr<VecTypeHintAttr>()) {
|
|
539 QualType HintQTy = A->getTypeHint();
|
|
540 const ExtVectorType *HintEltQTy = HintQTy->getAs<ExtVectorType>();
|
|
541 bool IsSignedInteger =
|
|
542 HintQTy->isSignedIntegerType() ||
|
|
543 (HintEltQTy && HintEltQTy->getElementType()->isSignedIntegerType());
|
|
544 llvm::Metadata *AttrMDArgs[] = {
|
|
545 llvm::ConstantAsMetadata::get(llvm::UndefValue::get(
|
|
546 CGM.getTypes().ConvertType(A->getTypeHint()))),
|
|
547 llvm::ConstantAsMetadata::get(llvm::ConstantInt::get(
|
|
548 llvm::IntegerType::get(Context, 32),
|
|
549 llvm::APInt(32, (uint64_t)(IsSignedInteger ? 1 : 0))))};
|
|
550 Fn->setMetadata("vec_type_hint", llvm::MDNode::get(Context, AttrMDArgs));
|
|
551 }
|
|
552
|
|
553 if (const WorkGroupSizeHintAttr *A = FD->getAttr<WorkGroupSizeHintAttr>()) {
|
|
554 llvm::Metadata *AttrMDArgs[] = {
|
|
555 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
|
|
556 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
|
|
557 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
|
|
558 Fn->setMetadata("work_group_size_hint", llvm::MDNode::get(Context, AttrMDArgs));
|
|
559 }
|
|
560
|
|
561 if (const ReqdWorkGroupSizeAttr *A = FD->getAttr<ReqdWorkGroupSizeAttr>()) {
|
|
562 llvm::Metadata *AttrMDArgs[] = {
|
|
563 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getXDim())),
|
|
564 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getYDim())),
|
|
565 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getZDim()))};
|
|
566 Fn->setMetadata("reqd_work_group_size", llvm::MDNode::get(Context, AttrMDArgs));
|
|
567 }
|
|
568
|
|
569 if (const OpenCLIntelReqdSubGroupSizeAttr *A =
|
|
570 FD->getAttr<OpenCLIntelReqdSubGroupSizeAttr>()) {
|
|
571 llvm::Metadata *AttrMDArgs[] = {
|
|
572 llvm::ConstantAsMetadata::get(Builder.getInt32(A->getSubGroupSize()))};
|
|
573 Fn->setMetadata("intel_reqd_sub_group_size",
|
|
574 llvm::MDNode::get(Context, AttrMDArgs));
|
|
575 }
|
|
576 }
|
|
577
|
|
578 /// Determine whether the function F ends with a return stmt.
|
|
579 static bool endsWithReturn(const Decl* F) {
|
|
580 const Stmt *Body = nullptr;
|
|
581 if (auto *FD = dyn_cast_or_null<FunctionDecl>(F))
|
|
582 Body = FD->getBody();
|
|
583 else if (auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(F))
|
|
584 Body = OMD->getBody();
|
|
585
|
|
586 if (auto *CS = dyn_cast_or_null<CompoundStmt>(Body)) {
|
|
587 auto LastStmt = CS->body_rbegin();
|
|
588 if (LastStmt != CS->body_rend())
|
|
589 return isa<ReturnStmt>(*LastStmt);
|
|
590 }
|
|
591 return false;
|
|
592 }
|
|
593
|
|
594 void CodeGenFunction::markAsIgnoreThreadCheckingAtRuntime(llvm::Function *Fn) {
|
|
595 if (SanOpts.has(SanitizerKind::Thread)) {
|
|
596 Fn->addFnAttr("sanitize_thread_no_checking_at_run_time");
|
|
597 Fn->removeFnAttr(llvm::Attribute::SanitizeThread);
|
|
598 }
|
|
599 }
|
|
600
|
|
601 /// Check if the return value of this function requires sanitization.
|
|
602 bool CodeGenFunction::requiresReturnValueCheck() const {
|
|
603 return requiresReturnValueNullabilityCheck() ||
|
|
604 (SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) && CurCodeDecl &&
|
|
605 CurCodeDecl->getAttr<ReturnsNonNullAttr>());
|
|
606 }
|
|
607
|
|
608 static bool matchesStlAllocatorFn(const Decl *D, const ASTContext &Ctx) {
|
|
609 auto *MD = dyn_cast_or_null<CXXMethodDecl>(D);
|
|
610 if (!MD || !MD->getDeclName().getAsIdentifierInfo() ||
|
|
611 !MD->getDeclName().getAsIdentifierInfo()->isStr("allocate") ||
|
|
612 (MD->getNumParams() != 1 && MD->getNumParams() != 2))
|
|
613 return false;
|
|
614
|
|
615 if (MD->parameters()[0]->getType().getCanonicalType() != Ctx.getSizeType())
|
|
616 return false;
|
|
617
|
|
618 if (MD->getNumParams() == 2) {
|
|
619 auto *PT = MD->parameters()[1]->getType()->getAs<PointerType>();
|
|
620 if (!PT || !PT->isVoidPointerType() ||
|
|
621 !PT->getPointeeType().isConstQualified())
|
|
622 return false;
|
|
623 }
|
|
624
|
|
625 return true;
|
|
626 }
|
|
627
|
|
628 /// Return the UBSan prologue signature for \p FD if one is available.
|
|
629 static llvm::Constant *getPrologueSignature(CodeGenModule &CGM,
|
|
630 const FunctionDecl *FD) {
|
|
631 if (const auto *MD = dyn_cast<CXXMethodDecl>(FD))
|
|
632 if (!MD->isStatic())
|
|
633 return nullptr;
|
|
634 return CGM.getTargetCodeGenInfo().getUBSanFunctionSignature(CGM);
|
|
635 }
|
|
636
|
|
637 void CodeGenFunction::StartFunction(GlobalDecl GD, QualType RetTy,
|
|
638 llvm::Function *Fn,
|
|
639 const CGFunctionInfo &FnInfo,
|
|
640 const FunctionArgList &Args,
|
|
641 SourceLocation Loc,
|
|
642 SourceLocation StartLoc) {
|
|
643 assert(!CurFn &&
|
|
644 "Do not use a CodeGenFunction object for more than one function");
|
|
645
|
|
646 const Decl *D = GD.getDecl();
|
|
647
|
|
648 DidCallStackSave = false;
|
|
649 CurCodeDecl = D;
|
|
650 if (const auto *FD = dyn_cast_or_null<FunctionDecl>(D))
|
|
651 if (FD->usesSEHTry())
|
|
652 CurSEHParent = FD;
|
|
653 CurFuncDecl = (D ? D->getNonClosureContext() : nullptr);
|
|
654 FnRetTy = RetTy;
|
|
655 CurFn = Fn;
|
|
656 CurFnInfo = &FnInfo;
|
|
657 assert(CurFn->isDeclaration() && "Function already has body?");
|
|
658
|
|
659 // If this function has been blacklisted for any of the enabled sanitizers,
|
|
660 // disable the sanitizer for the function.
|
|
661 do {
|
|
662 #define SANITIZER(NAME, ID) \
|
|
663 if (SanOpts.empty()) \
|
|
664 break; \
|
|
665 if (SanOpts.has(SanitizerKind::ID)) \
|
|
666 if (CGM.isInSanitizerBlacklist(SanitizerKind::ID, Fn, Loc)) \
|
|
667 SanOpts.set(SanitizerKind::ID, false);
|
|
668
|
|
669 #include "clang/Basic/Sanitizers.def"
|
|
670 #undef SANITIZER
|
|
671 } while (0);
|
|
672
|
|
673 if (D) {
|
|
674 // Apply the no_sanitize* attributes to SanOpts.
|
|
675 for (auto Attr : D->specific_attrs<NoSanitizeAttr>()) {
|
|
676 SanitizerMask mask = Attr->getMask();
|
|
677 SanOpts.Mask &= ~mask;
|
|
678 if (mask & SanitizerKind::Address)
|
|
679 SanOpts.set(SanitizerKind::KernelAddress, false);
|
|
680 if (mask & SanitizerKind::KernelAddress)
|
|
681 SanOpts.set(SanitizerKind::Address, false);
|
|
682 if (mask & SanitizerKind::HWAddress)
|
|
683 SanOpts.set(SanitizerKind::KernelHWAddress, false);
|
|
684 if (mask & SanitizerKind::KernelHWAddress)
|
|
685 SanOpts.set(SanitizerKind::HWAddress, false);
|
|
686 }
|
|
687 }
|
|
688
|
|
689 // Apply sanitizer attributes to the function.
|
|
690 if (SanOpts.hasOneOf(SanitizerKind::Address | SanitizerKind::KernelAddress))
|
|
691 Fn->addFnAttr(llvm::Attribute::SanitizeAddress);
|
|
692 if (SanOpts.hasOneOf(SanitizerKind::HWAddress | SanitizerKind::KernelHWAddress))
|
|
693 Fn->addFnAttr(llvm::Attribute::SanitizeHWAddress);
|
|
694 if (SanOpts.has(SanitizerKind::MemTag))
|
|
695 Fn->addFnAttr(llvm::Attribute::SanitizeMemTag);
|
|
696 if (SanOpts.has(SanitizerKind::Thread))
|
|
697 Fn->addFnAttr(llvm::Attribute::SanitizeThread);
|
|
698 if (SanOpts.hasOneOf(SanitizerKind::Memory | SanitizerKind::KernelMemory))
|
|
699 Fn->addFnAttr(llvm::Attribute::SanitizeMemory);
|
|
700 if (SanOpts.has(SanitizerKind::SafeStack))
|
|
701 Fn->addFnAttr(llvm::Attribute::SafeStack);
|
|
702 if (SanOpts.has(SanitizerKind::ShadowCallStack))
|
|
703 Fn->addFnAttr(llvm::Attribute::ShadowCallStack);
|
|
704
|
|
705 // Apply fuzzing attribute to the function.
|
|
706 if (SanOpts.hasOneOf(SanitizerKind::Fuzzer | SanitizerKind::FuzzerNoLink))
|
|
707 Fn->addFnAttr(llvm::Attribute::OptForFuzzing);
|
|
708
|
|
709 // Ignore TSan memory acesses from within ObjC/ObjC++ dealloc, initialize,
|
|
710 // .cxx_destruct, __destroy_helper_block_ and all of their calees at run time.
|
|
711 if (SanOpts.has(SanitizerKind::Thread)) {
|
|
712 if (const auto *OMD = dyn_cast_or_null<ObjCMethodDecl>(D)) {
|
|
713 IdentifierInfo *II = OMD->getSelector().getIdentifierInfoForSlot(0);
|
|
714 if (OMD->getMethodFamily() == OMF_dealloc ||
|
|
715 OMD->getMethodFamily() == OMF_initialize ||
|
|
716 (OMD->getSelector().isUnarySelector() && II->isStr(".cxx_destruct"))) {
|
|
717 markAsIgnoreThreadCheckingAtRuntime(Fn);
|
|
718 }
|
|
719 }
|
|
720 }
|
|
721
|
|
722 // Ignore unrelated casts in STL allocate() since the allocator must cast
|
|
723 // from void* to T* before object initialization completes. Don't match on the
|
|
724 // namespace because not all allocators are in std::
|
|
725 if (D && SanOpts.has(SanitizerKind::CFIUnrelatedCast)) {
|
|
726 if (matchesStlAllocatorFn(D, getContext()))
|
|
727 SanOpts.Mask &= ~SanitizerKind::CFIUnrelatedCast;
|
|
728 }
|
|
729
|
|
730 // Ignore null checks in coroutine functions since the coroutines passes
|
|
731 // are not aware of how to move the extra UBSan instructions across the split
|
|
732 // coroutine boundaries.
|
|
733 if (D && SanOpts.has(SanitizerKind::Null))
|
|
734 if (const auto *FD = dyn_cast<FunctionDecl>(D))
|
|
735 if (FD->getBody() &&
|
|
736 FD->getBody()->getStmtClass() == Stmt::CoroutineBodyStmtClass)
|
|
737 SanOpts.Mask &= ~SanitizerKind::Null;
|
|
738
|
173
|
739 // Apply xray attributes to the function (as a string, for now)
|
|
740 if (const auto *XRayAttr = D ? D->getAttr<XRayInstrumentAttr>() : nullptr) {
|
|
741 if (CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
|
|
742 XRayInstrKind::FunctionEntry) ||
|
|
743 CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
|
|
744 XRayInstrKind::FunctionExit)) {
|
|
745 if (XRayAttr->alwaysXRayInstrument() && ShouldXRayInstrumentFunction())
|
|
746 Fn->addFnAttr("function-instrument", "xray-always");
|
|
747 if (XRayAttr->neverXRayInstrument())
|
|
748 Fn->addFnAttr("function-instrument", "xray-never");
|
|
749 if (const auto *LogArgs = D->getAttr<XRayLogArgsAttr>())
|
|
750 if (ShouldXRayInstrumentFunction())
|
|
751 Fn->addFnAttr("xray-log-args",
|
|
752 llvm::utostr(LogArgs->getArgumentCount()));
|
150
|
753 }
|
173
|
754 } else {
|
|
755 if (ShouldXRayInstrumentFunction() && !CGM.imbueXRayAttrs(Fn, Loc))
|
|
756 Fn->addFnAttr(
|
|
757 "xray-instruction-threshold",
|
|
758 llvm::itostr(CGM.getCodeGenOpts().XRayInstructionThreshold));
|
|
759 }
|
150
|
760
|
173
|
761 if (ShouldXRayInstrumentFunction()) {
|
|
762 if (CGM.getCodeGenOpts().XRayIgnoreLoops)
|
|
763 Fn->addFnAttr("xray-ignore-loops");
|
150
|
764
|
173
|
765 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
|
|
766 XRayInstrKind::FunctionExit))
|
|
767 Fn->addFnAttr("xray-skip-exit");
|
150
|
768
|
173
|
769 if (!CGM.getCodeGenOpts().XRayInstrumentationBundle.has(
|
|
770 XRayInstrKind::FunctionEntry))
|
|
771 Fn->addFnAttr("xray-skip-entry");
|
|
772 }
|
150
|
773
|
173
|
774 unsigned Count, Offset;
|
|
775 if (const auto *Attr =
|
|
776 D ? D->getAttr<PatchableFunctionEntryAttr>() : nullptr) {
|
|
777 Count = Attr->getCount();
|
|
778 Offset = Attr->getOffset();
|
|
779 } else {
|
|
780 Count = CGM.getCodeGenOpts().PatchableFunctionEntryCount;
|
|
781 Offset = CGM.getCodeGenOpts().PatchableFunctionEntryOffset;
|
|
782 }
|
|
783 if (Count && Offset <= Count) {
|
|
784 Fn->addFnAttr("patchable-function-entry", std::to_string(Count - Offset));
|
|
785 if (Offset)
|
|
786 Fn->addFnAttr("patchable-function-prefix", std::to_string(Offset));
|
150
|
787 }
|
|
788
|
|
789 // Add no-jump-tables value.
|
|
790 Fn->addFnAttr("no-jump-tables",
|
|
791 llvm::toStringRef(CGM.getCodeGenOpts().NoUseJumpTables));
|
|
792
|
|
793 // Add no-inline-line-tables value.
|
|
794 if (CGM.getCodeGenOpts().NoInlineLineTables)
|
|
795 Fn->addFnAttr("no-inline-line-tables");
|
|
796
|
|
797 // Add profile-sample-accurate value.
|
|
798 if (CGM.getCodeGenOpts().ProfileSampleAccurate)
|
|
799 Fn->addFnAttr("profile-sample-accurate");
|
|
800
|
|
801 if (D && D->hasAttr<CFICanonicalJumpTableAttr>())
|
|
802 Fn->addFnAttr("cfi-canonical-jump-table");
|
|
803
|
|
804 if (getLangOpts().OpenCL) {
|
|
805 // Add metadata for a kernel function.
|
|
806 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
|
|
807 EmitOpenCLKernelMetadata(FD, Fn);
|
|
808 }
|
|
809
|
|
810 // If we are checking function types, emit a function type signature as
|
|
811 // prologue data.
|
|
812 if (getLangOpts().CPlusPlus && SanOpts.has(SanitizerKind::Function)) {
|
|
813 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
|
|
814 if (llvm::Constant *PrologueSig = getPrologueSignature(CGM, FD)) {
|
|
815 // Remove any (C++17) exception specifications, to allow calling e.g. a
|
|
816 // noexcept function through a non-noexcept pointer.
|
|
817 auto ProtoTy =
|
|
818 getContext().getFunctionTypeWithExceptionSpec(FD->getType(),
|
|
819 EST_None);
|
|
820 llvm::Constant *FTRTTIConst =
|
|
821 CGM.GetAddrOfRTTIDescriptor(ProtoTy, /*ForEH=*/true);
|
|
822 llvm::Constant *FTRTTIConstEncoded =
|
|
823 EncodeAddrForUseInPrologue(Fn, FTRTTIConst);
|
|
824 llvm::Constant *PrologueStructElems[] = {PrologueSig,
|
|
825 FTRTTIConstEncoded};
|
|
826 llvm::Constant *PrologueStructConst =
|
|
827 llvm::ConstantStruct::getAnon(PrologueStructElems, /*Packed=*/true);
|
|
828 Fn->setPrologueData(PrologueStructConst);
|
|
829 }
|
|
830 }
|
|
831 }
|
|
832
|
|
833 // If we're checking nullability, we need to know whether we can check the
|
|
834 // return value. Initialize the flag to 'true' and refine it in EmitParmDecl.
|
|
835 if (SanOpts.has(SanitizerKind::NullabilityReturn)) {
|
|
836 auto Nullability = FnRetTy->getNullability(getContext());
|
|
837 if (Nullability && *Nullability == NullabilityKind::NonNull) {
|
|
838 if (!(SanOpts.has(SanitizerKind::ReturnsNonnullAttribute) &&
|
|
839 CurCodeDecl && CurCodeDecl->getAttr<ReturnsNonNullAttr>()))
|
|
840 RetValNullabilityPrecondition =
|
|
841 llvm::ConstantInt::getTrue(getLLVMContext());
|
|
842 }
|
|
843 }
|
|
844
|
|
845 // If we're in C++ mode and the function name is "main", it is guaranteed
|
|
846 // to be norecurse by the standard (3.6.1.3 "The function main shall not be
|
|
847 // used within a program").
|
173
|
848 //
|
|
849 // OpenCL C 2.0 v2.2-11 s6.9.i:
|
|
850 // Recursion is not supported.
|
|
851 //
|
|
852 // SYCL v1.2.1 s3.10:
|
|
853 // kernels cannot include RTTI information, exception classes,
|
|
854 // recursive code, virtual functions or make use of C++ libraries that
|
|
855 // are not compiled for the device.
|
|
856 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
|
|
857 if ((getLangOpts().CPlusPlus && FD->isMain()) || getLangOpts().OpenCL ||
|
|
858 getLangOpts().SYCLIsDevice ||
|
|
859 (getLangOpts().CUDA && FD->hasAttr<CUDAGlobalAttr>()))
|
|
860 Fn->addFnAttr(llvm::Attribute::NoRecurse);
|
|
861 }
|
150
|
862
|
173
|
863 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D)) {
|
|
864 Builder.setIsFPConstrained(FD->usesFPIntrin());
|
150
|
865 if (FD->usesFPIntrin())
|
|
866 Fn->addFnAttr(llvm::Attribute::StrictFP);
|
173
|
867 }
|
150
|
868
|
|
869 // If a custom alignment is used, force realigning to this alignment on
|
|
870 // any main function which certainly will need it.
|
|
871 if (const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(D))
|
|
872 if ((FD->isMain() || FD->isMSVCRTEntryPoint()) &&
|
|
873 CGM.getCodeGenOpts().StackAlignment)
|
|
874 Fn->addFnAttr("stackrealign");
|
|
875
|
|
876 llvm::BasicBlock *EntryBB = createBasicBlock("entry", CurFn);
|
|
877
|
|
878 // Create a marker to make it easy to insert allocas into the entryblock
|
|
879 // later. Don't create this with the builder, because we don't want it
|
|
880 // folded.
|
|
881 llvm::Value *Undef = llvm::UndefValue::get(Int32Ty);
|
|
882 AllocaInsertPt = new llvm::BitCastInst(Undef, Int32Ty, "allocapt", EntryBB);
|
|
883
|
|
884 ReturnBlock = getJumpDestInCurrentScope("return");
|
|
885
|
|
886 Builder.SetInsertPoint(EntryBB);
|
|
887
|
|
888 // If we're checking the return value, allocate space for a pointer to a
|
|
889 // precise source location of the checked return statement.
|
|
890 if (requiresReturnValueCheck()) {
|
|
891 ReturnLocation = CreateDefaultAlignTempAlloca(Int8PtrTy, "return.sloc.ptr");
|
|
892 InitTempAlloca(ReturnLocation, llvm::ConstantPointerNull::get(Int8PtrTy));
|
|
893 }
|
|
894
|
|
895 // Emit subprogram debug descriptor.
|
|
896 if (CGDebugInfo *DI = getDebugInfo()) {
|
|
897 // Reconstruct the type from the argument list so that implicit parameters,
|
|
898 // such as 'this' and 'vtt', show up in the debug info. Preserve the calling
|
|
899 // convention.
|
|
900 CallingConv CC = CallingConv::CC_C;
|
|
901 if (auto *FD = dyn_cast_or_null<FunctionDecl>(D))
|
|
902 if (const auto *SrcFnTy = FD->getType()->getAs<FunctionType>())
|
|
903 CC = SrcFnTy->getCallConv();
|
|
904 SmallVector<QualType, 16> ArgTypes;
|
|
905 for (const VarDecl *VD : Args)
|
|
906 ArgTypes.push_back(VD->getType());
|
|
907 QualType FnType = getContext().getFunctionType(
|
|
908 RetTy, ArgTypes, FunctionProtoType::ExtProtoInfo(CC));
|
|
909 DI->EmitFunctionStart(GD, Loc, StartLoc, FnType, CurFn, CurFuncIsThunk,
|
|
910 Builder);
|
|
911 }
|
|
912
|
|
913 if (ShouldInstrumentFunction()) {
|
|
914 if (CGM.getCodeGenOpts().InstrumentFunctions)
|
|
915 CurFn->addFnAttr("instrument-function-entry", "__cyg_profile_func_enter");
|
|
916 if (CGM.getCodeGenOpts().InstrumentFunctionsAfterInlining)
|
|
917 CurFn->addFnAttr("instrument-function-entry-inlined",
|
|
918 "__cyg_profile_func_enter");
|
|
919 if (CGM.getCodeGenOpts().InstrumentFunctionEntryBare)
|
|
920 CurFn->addFnAttr("instrument-function-entry-inlined",
|
|
921 "__cyg_profile_func_enter_bare");
|
|
922 }
|
|
923
|
|
924 // Since emitting the mcount call here impacts optimizations such as function
|
|
925 // inlining, we just add an attribute to insert a mcount call in backend.
|
|
926 // The attribute "counting-function" is set to mcount function name which is
|
|
927 // architecture dependent.
|
|
928 if (CGM.getCodeGenOpts().InstrumentForProfiling) {
|
|
929 // Calls to fentry/mcount should not be generated if function has
|
|
930 // the no_instrument_function attribute.
|
|
931 if (!CurFuncDecl || !CurFuncDecl->hasAttr<NoInstrumentFunctionAttr>()) {
|
|
932 if (CGM.getCodeGenOpts().CallFEntry)
|
|
933 Fn->addFnAttr("fentry-call", "true");
|
|
934 else {
|
|
935 Fn->addFnAttr("instrument-function-entry-inlined",
|
|
936 getTarget().getMCountName());
|
|
937 }
|
|
938 if (CGM.getCodeGenOpts().MNopMCount) {
|
|
939 if (!CGM.getCodeGenOpts().CallFEntry)
|
|
940 CGM.getDiags().Report(diag::err_opt_not_valid_without_opt)
|
|
941 << "-mnop-mcount" << "-mfentry";
|
|
942 Fn->addFnAttr("mnop-mcount");
|
|
943 }
|
|
944
|
|
945 if (CGM.getCodeGenOpts().RecordMCount) {
|
|
946 if (!CGM.getCodeGenOpts().CallFEntry)
|
|
947 CGM.getDiags().Report(diag::err_opt_not_valid_without_opt)
|
|
948 << "-mrecord-mcount" << "-mfentry";
|
|
949 Fn->addFnAttr("mrecord-mcount");
|
|
950 }
|
|
951 }
|
|
952 }
|
|
953
|
|
954 if (CGM.getCodeGenOpts().PackedStack) {
|
|
955 if (getContext().getTargetInfo().getTriple().getArch() !=
|
|
956 llvm::Triple::systemz)
|
|
957 CGM.getDiags().Report(diag::err_opt_not_valid_on_target)
|
|
958 << "-mpacked-stack";
|
|
959 Fn->addFnAttr("packed-stack");
|
|
960 }
|
|
961
|
|
962 if (RetTy->isVoidType()) {
|
|
963 // Void type; nothing to return.
|
|
964 ReturnValue = Address::invalid();
|
|
965
|
|
966 // Count the implicit return.
|
|
967 if (!endsWithReturn(D))
|
|
968 ++NumReturnExprs;
|
|
969 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::Indirect) {
|
|
970 // Indirect return; emit returned value directly into sret slot.
|
|
971 // This reduces code size, and affects correctness in C++.
|
|
972 auto AI = CurFn->arg_begin();
|
|
973 if (CurFnInfo->getReturnInfo().isSRetAfterThis())
|
|
974 ++AI;
|
|
975 ReturnValue = Address(&*AI, CurFnInfo->getReturnInfo().getIndirectAlign());
|
|
976 if (!CurFnInfo->getReturnInfo().getIndirectByVal()) {
|
|
977 ReturnValuePointer =
|
|
978 CreateDefaultAlignTempAlloca(Int8PtrTy, "result.ptr");
|
|
979 Builder.CreateStore(Builder.CreatePointerBitCastOrAddrSpaceCast(
|
|
980 ReturnValue.getPointer(), Int8PtrTy),
|
|
981 ReturnValuePointer);
|
|
982 }
|
|
983 } else if (CurFnInfo->getReturnInfo().getKind() == ABIArgInfo::InAlloca &&
|
|
984 !hasScalarEvaluationKind(CurFnInfo->getReturnType())) {
|
|
985 // Load the sret pointer from the argument struct and return into that.
|
|
986 unsigned Idx = CurFnInfo->getReturnInfo().getInAllocaFieldIndex();
|
|
987 llvm::Function::arg_iterator EI = CurFn->arg_end();
|
|
988 --EI;
|
|
989 llvm::Value *Addr = Builder.CreateStructGEP(nullptr, &*EI, Idx);
|
|
990 ReturnValuePointer = Address(Addr, getPointerAlign());
|
|
991 Addr = Builder.CreateAlignedLoad(Addr, getPointerAlign(), "agg.result");
|
173
|
992 ReturnValue = Address(Addr, CGM.getNaturalTypeAlignment(RetTy));
|
150
|
993 } else {
|
|
994 ReturnValue = CreateIRTemp(RetTy, "retval");
|
|
995
|
|
996 // Tell the epilog emitter to autorelease the result. We do this
|
|
997 // now so that various specialized functions can suppress it
|
|
998 // during their IR-generation.
|
|
999 if (getLangOpts().ObjCAutoRefCount &&
|
|
1000 !CurFnInfo->isReturnsRetained() &&
|
|
1001 RetTy->isObjCRetainableType())
|
|
1002 AutoreleaseResult = true;
|
|
1003 }
|
|
1004
|
|
1005 EmitStartEHSpec(CurCodeDecl);
|
|
1006
|
|
1007 PrologueCleanupDepth = EHStack.stable_begin();
|
|
1008
|
|
1009 // Emit OpenMP specific initialization of the device functions.
|
|
1010 if (getLangOpts().OpenMP && CurCodeDecl)
|
|
1011 CGM.getOpenMPRuntime().emitFunctionProlog(*this, CurCodeDecl);
|
|
1012
|
|
1013 EmitFunctionProlog(*CurFnInfo, CurFn, Args);
|
|
1014
|
|
1015 if (D && isa<CXXMethodDecl>(D) && cast<CXXMethodDecl>(D)->isInstance()) {
|
|
1016 CGM.getCXXABI().EmitInstanceFunctionProlog(*this);
|
|
1017 const CXXMethodDecl *MD = cast<CXXMethodDecl>(D);
|
|
1018 if (MD->getParent()->isLambda() &&
|
|
1019 MD->getOverloadedOperator() == OO_Call) {
|
|
1020 // We're in a lambda; figure out the captures.
|
|
1021 MD->getParent()->getCaptureFields(LambdaCaptureFields,
|
|
1022 LambdaThisCaptureField);
|
|
1023 if (LambdaThisCaptureField) {
|
|
1024 // If the lambda captures the object referred to by '*this' - either by
|
|
1025 // value or by reference, make sure CXXThisValue points to the correct
|
|
1026 // object.
|
|
1027
|
|
1028 // Get the lvalue for the field (which is a copy of the enclosing object
|
|
1029 // or contains the address of the enclosing object).
|
|
1030 LValue ThisFieldLValue = EmitLValueForLambdaField(LambdaThisCaptureField);
|
|
1031 if (!LambdaThisCaptureField->getType()->isPointerType()) {
|
|
1032 // If the enclosing object was captured by value, just use its address.
|
|
1033 CXXThisValue = ThisFieldLValue.getAddress(*this).getPointer();
|
|
1034 } else {
|
|
1035 // Load the lvalue pointed to by the field, since '*this' was captured
|
|
1036 // by reference.
|
|
1037 CXXThisValue =
|
|
1038 EmitLoadOfLValue(ThisFieldLValue, SourceLocation()).getScalarVal();
|
|
1039 }
|
|
1040 }
|
|
1041 for (auto *FD : MD->getParent()->fields()) {
|
|
1042 if (FD->hasCapturedVLAType()) {
|
|
1043 auto *ExprArg = EmitLoadOfLValue(EmitLValueForLambdaField(FD),
|
|
1044 SourceLocation()).getScalarVal();
|
|
1045 auto VAT = FD->getCapturedVLAType();
|
|
1046 VLASizeMap[VAT->getSizeExpr()] = ExprArg;
|
|
1047 }
|
|
1048 }
|
|
1049 } else {
|
|
1050 // Not in a lambda; just use 'this' from the method.
|
|
1051 // FIXME: Should we generate a new load for each use of 'this'? The
|
|
1052 // fast register allocator would be happier...
|
|
1053 CXXThisValue = CXXABIThisValue;
|
|
1054 }
|
|
1055
|
|
1056 // Check the 'this' pointer once per function, if it's available.
|
|
1057 if (CXXABIThisValue) {
|
|
1058 SanitizerSet SkippedChecks;
|
|
1059 SkippedChecks.set(SanitizerKind::ObjectSize, true);
|
|
1060 QualType ThisTy = MD->getThisType();
|
|
1061
|
|
1062 // If this is the call operator of a lambda with no capture-default, it
|
|
1063 // may have a static invoker function, which may call this operator with
|
|
1064 // a null 'this' pointer.
|
|
1065 if (isLambdaCallOperator(MD) &&
|
|
1066 MD->getParent()->getLambdaCaptureDefault() == LCD_None)
|
|
1067 SkippedChecks.set(SanitizerKind::Null, true);
|
|
1068
|
|
1069 EmitTypeCheck(isa<CXXConstructorDecl>(MD) ? TCK_ConstructorCall
|
|
1070 : TCK_MemberCall,
|
|
1071 Loc, CXXABIThisValue, ThisTy,
|
|
1072 getContext().getTypeAlignInChars(ThisTy->getPointeeType()),
|
|
1073 SkippedChecks);
|
|
1074 }
|
|
1075 }
|
|
1076
|
|
1077 // If any of the arguments have a variably modified type, make sure to
|
|
1078 // emit the type size.
|
|
1079 for (FunctionArgList::const_iterator i = Args.begin(), e = Args.end();
|
|
1080 i != e; ++i) {
|
|
1081 const VarDecl *VD = *i;
|
|
1082
|
|
1083 // Dig out the type as written from ParmVarDecls; it's unclear whether
|
|
1084 // the standard (C99 6.9.1p10) requires this, but we're following the
|
|
1085 // precedent set by gcc.
|
|
1086 QualType Ty;
|
|
1087 if (const ParmVarDecl *PVD = dyn_cast<ParmVarDecl>(VD))
|
|
1088 Ty = PVD->getOriginalType();
|
|
1089 else
|
|
1090 Ty = VD->getType();
|
|
1091
|
|
1092 if (Ty->isVariablyModifiedType())
|
|
1093 EmitVariablyModifiedType(Ty);
|
|
1094 }
|
|
1095 // Emit a location at the end of the prologue.
|
|
1096 if (CGDebugInfo *DI = getDebugInfo())
|
|
1097 DI->EmitLocation(Builder, StartLoc);
|
|
1098
|
|
1099 // TODO: Do we need to handle this in two places like we do with
|
|
1100 // target-features/target-cpu?
|
|
1101 if (CurFuncDecl)
|
|
1102 if (const auto *VecWidth = CurFuncDecl->getAttr<MinVectorWidthAttr>())
|
|
1103 LargestVectorWidth = VecWidth->getVectorWidth();
|
|
1104 }
|
|
1105
|
|
1106 void CodeGenFunction::EmitFunctionBody(const Stmt *Body) {
|
|
1107 incrementProfileCounter(Body);
|
|
1108 if (const CompoundStmt *S = dyn_cast<CompoundStmt>(Body))
|
|
1109 EmitCompoundStmtWithoutScope(*S);
|
|
1110 else
|
|
1111 EmitStmt(Body);
|
|
1112 }
|
|
1113
|
|
1114 /// When instrumenting to collect profile data, the counts for some blocks
|
|
1115 /// such as switch cases need to not include the fall-through counts, so
|
|
1116 /// emit a branch around the instrumentation code. When not instrumenting,
|
|
1117 /// this just calls EmitBlock().
|
|
1118 void CodeGenFunction::EmitBlockWithFallThrough(llvm::BasicBlock *BB,
|
|
1119 const Stmt *S) {
|
|
1120 llvm::BasicBlock *SkipCountBB = nullptr;
|
|
1121 if (HaveInsertPoint() && CGM.getCodeGenOpts().hasProfileClangInstr()) {
|
|
1122 // When instrumenting for profiling, the fallthrough to certain
|
|
1123 // statements needs to skip over the instrumentation code so that we
|
|
1124 // get an accurate count.
|
|
1125 SkipCountBB = createBasicBlock("skipcount");
|
|
1126 EmitBranch(SkipCountBB);
|
|
1127 }
|
|
1128 EmitBlock(BB);
|
|
1129 uint64_t CurrentCount = getCurrentProfileCount();
|
|
1130 incrementProfileCounter(S);
|
|
1131 setCurrentProfileCount(getCurrentProfileCount() + CurrentCount);
|
|
1132 if (SkipCountBB)
|
|
1133 EmitBlock(SkipCountBB);
|
|
1134 }
|
|
1135
|
|
1136 /// Tries to mark the given function nounwind based on the
|
|
1137 /// non-existence of any throwing calls within it. We believe this is
|
|
1138 /// lightweight enough to do at -O0.
|
|
1139 static void TryMarkNoThrow(llvm::Function *F) {
|
|
1140 // LLVM treats 'nounwind' on a function as part of the type, so we
|
|
1141 // can't do this on functions that can be overwritten.
|
|
1142 if (F->isInterposable()) return;
|
|
1143
|
|
1144 for (llvm::BasicBlock &BB : *F)
|
|
1145 for (llvm::Instruction &I : BB)
|
|
1146 if (I.mayThrow())
|
|
1147 return;
|
|
1148
|
|
1149 F->setDoesNotThrow();
|
|
1150 }
|
|
1151
|
|
1152 QualType CodeGenFunction::BuildFunctionArgList(GlobalDecl GD,
|
|
1153 FunctionArgList &Args) {
|
|
1154 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
|
|
1155 QualType ResTy = FD->getReturnType();
|
|
1156
|
|
1157 const CXXMethodDecl *MD = dyn_cast<CXXMethodDecl>(FD);
|
|
1158 if (MD && MD->isInstance()) {
|
|
1159 if (CGM.getCXXABI().HasThisReturn(GD))
|
|
1160 ResTy = MD->getThisType();
|
|
1161 else if (CGM.getCXXABI().hasMostDerivedReturn(GD))
|
|
1162 ResTy = CGM.getContext().VoidPtrTy;
|
|
1163 CGM.getCXXABI().buildThisParam(*this, Args);
|
|
1164 }
|
|
1165
|
|
1166 // The base version of an inheriting constructor whose constructed base is a
|
|
1167 // virtual base is not passed any arguments (because it doesn't actually call
|
|
1168 // the inherited constructor).
|
|
1169 bool PassedParams = true;
|
|
1170 if (const CXXConstructorDecl *CD = dyn_cast<CXXConstructorDecl>(FD))
|
|
1171 if (auto Inherited = CD->getInheritedConstructor())
|
|
1172 PassedParams =
|
|
1173 getTypes().inheritingCtorHasParams(Inherited, GD.getCtorType());
|
|
1174
|
|
1175 if (PassedParams) {
|
|
1176 for (auto *Param : FD->parameters()) {
|
|
1177 Args.push_back(Param);
|
|
1178 if (!Param->hasAttr<PassObjectSizeAttr>())
|
|
1179 continue;
|
|
1180
|
|
1181 auto *Implicit = ImplicitParamDecl::Create(
|
|
1182 getContext(), Param->getDeclContext(), Param->getLocation(),
|
|
1183 /*Id=*/nullptr, getContext().getSizeType(), ImplicitParamDecl::Other);
|
|
1184 SizeArguments[Param] = Implicit;
|
|
1185 Args.push_back(Implicit);
|
|
1186 }
|
|
1187 }
|
|
1188
|
|
1189 if (MD && (isa<CXXConstructorDecl>(MD) || isa<CXXDestructorDecl>(MD)))
|
|
1190 CGM.getCXXABI().addImplicitStructorParams(*this, ResTy, Args);
|
|
1191
|
|
1192 return ResTy;
|
|
1193 }
|
|
1194
|
|
1195 static bool
|
|
1196 shouldUseUndefinedBehaviorReturnOptimization(const FunctionDecl *FD,
|
|
1197 const ASTContext &Context) {
|
|
1198 QualType T = FD->getReturnType();
|
|
1199 // Avoid the optimization for functions that return a record type with a
|
|
1200 // trivial destructor or another trivially copyable type.
|
|
1201 if (const RecordType *RT = T.getCanonicalType()->getAs<RecordType>()) {
|
|
1202 if (const auto *ClassDecl = dyn_cast<CXXRecordDecl>(RT->getDecl()))
|
|
1203 return !ClassDecl->hasTrivialDestructor();
|
|
1204 }
|
|
1205 return !T.isTriviallyCopyableType(Context);
|
|
1206 }
|
|
1207
|
|
1208 void CodeGenFunction::GenerateCode(GlobalDecl GD, llvm::Function *Fn,
|
|
1209 const CGFunctionInfo &FnInfo) {
|
|
1210 const FunctionDecl *FD = cast<FunctionDecl>(GD.getDecl());
|
|
1211 CurGD = GD;
|
|
1212
|
|
1213 FunctionArgList Args;
|
|
1214 QualType ResTy = BuildFunctionArgList(GD, Args);
|
|
1215
|
|
1216 // Check if we should generate debug info for this function.
|
|
1217 if (FD->hasAttr<NoDebugAttr>())
|
|
1218 DebugInfo = nullptr; // disable debug info indefinitely for this function
|
|
1219
|
|
1220 // The function might not have a body if we're generating thunks for a
|
|
1221 // function declaration.
|
|
1222 SourceRange BodyRange;
|
|
1223 if (Stmt *Body = FD->getBody())
|
|
1224 BodyRange = Body->getSourceRange();
|
|
1225 else
|
|
1226 BodyRange = FD->getLocation();
|
|
1227 CurEHLocation = BodyRange.getEnd();
|
|
1228
|
|
1229 // Use the location of the start of the function to determine where
|
|
1230 // the function definition is located. By default use the location
|
|
1231 // of the declaration as the location for the subprogram. A function
|
|
1232 // may lack a declaration in the source code if it is created by code
|
|
1233 // gen. (examples: _GLOBAL__I_a, __cxx_global_array_dtor, thunk).
|
|
1234 SourceLocation Loc = FD->getLocation();
|
|
1235
|
|
1236 // If this is a function specialization then use the pattern body
|
|
1237 // as the location for the function.
|
|
1238 if (const FunctionDecl *SpecDecl = FD->getTemplateInstantiationPattern())
|
|
1239 if (SpecDecl->hasBody(SpecDecl))
|
|
1240 Loc = SpecDecl->getLocation();
|
|
1241
|
|
1242 Stmt *Body = FD->getBody();
|
|
1243
|
|
1244 // Initialize helper which will detect jumps which can cause invalid lifetime
|
|
1245 // markers.
|
|
1246 if (Body && ShouldEmitLifetimeMarkers)
|
|
1247 Bypasses.Init(Body);
|
|
1248
|
|
1249 // Emit the standard function prologue.
|
|
1250 StartFunction(GD, ResTy, Fn, FnInfo, Args, Loc, BodyRange.getBegin());
|
|
1251
|
|
1252 // Generate the body of the function.
|
|
1253 PGO.assignRegionCounters(GD, CurFn);
|
|
1254 if (isa<CXXDestructorDecl>(FD))
|
|
1255 EmitDestructorBody(Args);
|
|
1256 else if (isa<CXXConstructorDecl>(FD))
|
|
1257 EmitConstructorBody(Args);
|
|
1258 else if (getLangOpts().CUDA &&
|
|
1259 !getLangOpts().CUDAIsDevice &&
|
|
1260 FD->hasAttr<CUDAGlobalAttr>())
|
|
1261 CGM.getCUDARuntime().emitDeviceStub(*this, Args);
|
|
1262 else if (isa<CXXMethodDecl>(FD) &&
|
|
1263 cast<CXXMethodDecl>(FD)->isLambdaStaticInvoker()) {
|
|
1264 // The lambda static invoker function is special, because it forwards or
|
|
1265 // clones the body of the function call operator (but is actually static).
|
|
1266 EmitLambdaStaticInvokeBody(cast<CXXMethodDecl>(FD));
|
|
1267 } else if (FD->isDefaulted() && isa<CXXMethodDecl>(FD) &&
|
|
1268 (cast<CXXMethodDecl>(FD)->isCopyAssignmentOperator() ||
|
|
1269 cast<CXXMethodDecl>(FD)->isMoveAssignmentOperator())) {
|
|
1270 // Implicit copy-assignment gets the same special treatment as implicit
|
|
1271 // copy-constructors.
|
|
1272 emitImplicitAssignmentOperatorBody(Args);
|
|
1273 } else if (Body) {
|
|
1274 EmitFunctionBody(Body);
|
|
1275 } else
|
|
1276 llvm_unreachable("no definition for emitted function");
|
|
1277
|
|
1278 // C++11 [stmt.return]p2:
|
|
1279 // Flowing off the end of a function [...] results in undefined behavior in
|
|
1280 // a value-returning function.
|
|
1281 // C11 6.9.1p12:
|
|
1282 // If the '}' that terminates a function is reached, and the value of the
|
|
1283 // function call is used by the caller, the behavior is undefined.
|
|
1284 if (getLangOpts().CPlusPlus && !FD->hasImplicitReturnZero() && !SawAsmBlock &&
|
|
1285 !FD->getReturnType()->isVoidType() && Builder.GetInsertBlock()) {
|
|
1286 bool ShouldEmitUnreachable =
|
|
1287 CGM.getCodeGenOpts().StrictReturn ||
|
|
1288 shouldUseUndefinedBehaviorReturnOptimization(FD, getContext());
|
|
1289 if (SanOpts.has(SanitizerKind::Return)) {
|
|
1290 SanitizerScope SanScope(this);
|
|
1291 llvm::Value *IsFalse = Builder.getFalse();
|
|
1292 EmitCheck(std::make_pair(IsFalse, SanitizerKind::Return),
|
|
1293 SanitizerHandler::MissingReturn,
|
|
1294 EmitCheckSourceLocation(FD->getLocation()), None);
|
|
1295 } else if (ShouldEmitUnreachable) {
|
|
1296 if (CGM.getCodeGenOpts().OptimizationLevel == 0)
|
|
1297 EmitTrapCall(llvm::Intrinsic::trap);
|
|
1298 }
|
|
1299 if (SanOpts.has(SanitizerKind::Return) || ShouldEmitUnreachable) {
|
|
1300 Builder.CreateUnreachable();
|
|
1301 Builder.ClearInsertionPoint();
|
|
1302 }
|
|
1303 }
|
|
1304
|
|
1305 // Emit the standard function epilogue.
|
|
1306 FinishFunction(BodyRange.getEnd());
|
|
1307
|
|
1308 // If we haven't marked the function nothrow through other means, do
|
|
1309 // a quick pass now to see if we can.
|
|
1310 if (!CurFn->doesNotThrow())
|
|
1311 TryMarkNoThrow(CurFn);
|
|
1312 }
|
|
1313
|
|
1314 /// ContainsLabel - Return true if the statement contains a label in it. If
|
|
1315 /// this statement is not executed normally, it not containing a label means
|
|
1316 /// that we can just remove the code.
|
|
1317 bool CodeGenFunction::ContainsLabel(const Stmt *S, bool IgnoreCaseStmts) {
|
|
1318 // Null statement, not a label!
|
|
1319 if (!S) return false;
|
|
1320
|
|
1321 // If this is a label, we have to emit the code, consider something like:
|
|
1322 // if (0) { ... foo: bar(); } goto foo;
|
|
1323 //
|
|
1324 // TODO: If anyone cared, we could track __label__'s, since we know that you
|
|
1325 // can't jump to one from outside their declared region.
|
|
1326 if (isa<LabelStmt>(S))
|
|
1327 return true;
|
|
1328
|
|
1329 // If this is a case/default statement, and we haven't seen a switch, we have
|
|
1330 // to emit the code.
|
|
1331 if (isa<SwitchCase>(S) && !IgnoreCaseStmts)
|
|
1332 return true;
|
|
1333
|
|
1334 // If this is a switch statement, we want to ignore cases below it.
|
|
1335 if (isa<SwitchStmt>(S))
|
|
1336 IgnoreCaseStmts = true;
|
|
1337
|
|
1338 // Scan subexpressions for verboten labels.
|
|
1339 for (const Stmt *SubStmt : S->children())
|
|
1340 if (ContainsLabel(SubStmt, IgnoreCaseStmts))
|
|
1341 return true;
|
|
1342
|
|
1343 return false;
|
|
1344 }
|
|
1345
|
|
1346 /// containsBreak - Return true if the statement contains a break out of it.
|
|
1347 /// If the statement (recursively) contains a switch or loop with a break
|
|
1348 /// inside of it, this is fine.
|
|
1349 bool CodeGenFunction::containsBreak(const Stmt *S) {
|
|
1350 // Null statement, not a label!
|
|
1351 if (!S) return false;
|
|
1352
|
|
1353 // If this is a switch or loop that defines its own break scope, then we can
|
|
1354 // include it and anything inside of it.
|
|
1355 if (isa<SwitchStmt>(S) || isa<WhileStmt>(S) || isa<DoStmt>(S) ||
|
|
1356 isa<ForStmt>(S))
|
|
1357 return false;
|
|
1358
|
|
1359 if (isa<BreakStmt>(S))
|
|
1360 return true;
|
|
1361
|
|
1362 // Scan subexpressions for verboten breaks.
|
|
1363 for (const Stmt *SubStmt : S->children())
|
|
1364 if (containsBreak(SubStmt))
|
|
1365 return true;
|
|
1366
|
|
1367 return false;
|
|
1368 }
|
|
1369
|
|
1370 bool CodeGenFunction::mightAddDeclToScope(const Stmt *S) {
|
|
1371 if (!S) return false;
|
|
1372
|
|
1373 // Some statement kinds add a scope and thus never add a decl to the current
|
|
1374 // scope. Note, this list is longer than the list of statements that might
|
|
1375 // have an unscoped decl nested within them, but this way is conservatively
|
|
1376 // correct even if more statement kinds are added.
|
|
1377 if (isa<IfStmt>(S) || isa<SwitchStmt>(S) || isa<WhileStmt>(S) ||
|
|
1378 isa<DoStmt>(S) || isa<ForStmt>(S) || isa<CompoundStmt>(S) ||
|
|
1379 isa<CXXForRangeStmt>(S) || isa<CXXTryStmt>(S) ||
|
|
1380 isa<ObjCForCollectionStmt>(S) || isa<ObjCAtTryStmt>(S))
|
|
1381 return false;
|
|
1382
|
|
1383 if (isa<DeclStmt>(S))
|
|
1384 return true;
|
|
1385
|
|
1386 for (const Stmt *SubStmt : S->children())
|
|
1387 if (mightAddDeclToScope(SubStmt))
|
|
1388 return true;
|
|
1389
|
|
1390 return false;
|
|
1391 }
|
|
1392
|
|
1393 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
|
|
1394 /// to a constant, or if it does but contains a label, return false. If it
|
|
1395 /// constant folds return true and set the boolean result in Result.
|
|
1396 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
|
|
1397 bool &ResultBool,
|
|
1398 bool AllowLabels) {
|
|
1399 llvm::APSInt ResultInt;
|
|
1400 if (!ConstantFoldsToSimpleInteger(Cond, ResultInt, AllowLabels))
|
|
1401 return false;
|
|
1402
|
|
1403 ResultBool = ResultInt.getBoolValue();
|
|
1404 return true;
|
|
1405 }
|
|
1406
|
|
1407 /// ConstantFoldsToSimpleInteger - If the specified expression does not fold
|
|
1408 /// to a constant, or if it does but contains a label, return false. If it
|
|
1409 /// constant folds return true and set the folded value.
|
|
1410 bool CodeGenFunction::ConstantFoldsToSimpleInteger(const Expr *Cond,
|
|
1411 llvm::APSInt &ResultInt,
|
|
1412 bool AllowLabels) {
|
|
1413 // FIXME: Rename and handle conversion of other evaluatable things
|
|
1414 // to bool.
|
|
1415 Expr::EvalResult Result;
|
|
1416 if (!Cond->EvaluateAsInt(Result, getContext()))
|
|
1417 return false; // Not foldable, not integer or not fully evaluatable.
|
|
1418
|
|
1419 llvm::APSInt Int = Result.Val.getInt();
|
|
1420 if (!AllowLabels && CodeGenFunction::ContainsLabel(Cond))
|
|
1421 return false; // Contains a label.
|
|
1422
|
|
1423 ResultInt = Int;
|
|
1424 return true;
|
|
1425 }
|
|
1426
|
|
1427
|
|
1428
|
|
1429 /// EmitBranchOnBoolExpr - Emit a branch on a boolean condition (e.g. for an if
|
|
1430 /// statement) to the specified blocks. Based on the condition, this might try
|
|
1431 /// to simplify the codegen of the conditional based on the branch.
|
|
1432 ///
|
|
1433 void CodeGenFunction::EmitBranchOnBoolExpr(const Expr *Cond,
|
|
1434 llvm::BasicBlock *TrueBlock,
|
|
1435 llvm::BasicBlock *FalseBlock,
|
|
1436 uint64_t TrueCount) {
|
|
1437 Cond = Cond->IgnoreParens();
|
|
1438
|
|
1439 if (const BinaryOperator *CondBOp = dyn_cast<BinaryOperator>(Cond)) {
|
|
1440
|
|
1441 // Handle X && Y in a condition.
|
|
1442 if (CondBOp->getOpcode() == BO_LAnd) {
|
|
1443 // If we have "1 && X", simplify the code. "0 && X" would have constant
|
|
1444 // folded if the case was simple enough.
|
|
1445 bool ConstantBool = false;
|
|
1446 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
|
|
1447 ConstantBool) {
|
|
1448 // br(1 && X) -> br(X).
|
|
1449 incrementProfileCounter(CondBOp);
|
|
1450 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
|
|
1451 TrueCount);
|
|
1452 }
|
|
1453
|
|
1454 // If we have "X && 1", simplify the code to use an uncond branch.
|
|
1455 // "X && 0" would have been constant folded to 0.
|
|
1456 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
|
|
1457 ConstantBool) {
|
|
1458 // br(X && 1) -> br(X).
|
|
1459 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
|
|
1460 TrueCount);
|
|
1461 }
|
|
1462
|
|
1463 // Emit the LHS as a conditional. If the LHS conditional is false, we
|
|
1464 // want to jump to the FalseBlock.
|
|
1465 llvm::BasicBlock *LHSTrue = createBasicBlock("land.lhs.true");
|
|
1466 // The counter tells us how often we evaluate RHS, and all of TrueCount
|
|
1467 // can be propagated to that branch.
|
|
1468 uint64_t RHSCount = getProfileCount(CondBOp->getRHS());
|
|
1469
|
|
1470 ConditionalEvaluation eval(*this);
|
|
1471 {
|
|
1472 ApplyDebugLocation DL(*this, Cond);
|
|
1473 EmitBranchOnBoolExpr(CondBOp->getLHS(), LHSTrue, FalseBlock, RHSCount);
|
|
1474 EmitBlock(LHSTrue);
|
|
1475 }
|
|
1476
|
|
1477 incrementProfileCounter(CondBOp);
|
|
1478 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
|
|
1479
|
|
1480 // Any temporaries created here are conditional.
|
|
1481 eval.begin(*this);
|
|
1482 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, TrueCount);
|
|
1483 eval.end(*this);
|
|
1484
|
|
1485 return;
|
|
1486 }
|
|
1487
|
|
1488 if (CondBOp->getOpcode() == BO_LOr) {
|
|
1489 // If we have "0 || X", simplify the code. "1 || X" would have constant
|
|
1490 // folded if the case was simple enough.
|
|
1491 bool ConstantBool = false;
|
|
1492 if (ConstantFoldsToSimpleInteger(CondBOp->getLHS(), ConstantBool) &&
|
|
1493 !ConstantBool) {
|
|
1494 // br(0 || X) -> br(X).
|
|
1495 incrementProfileCounter(CondBOp);
|
|
1496 return EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock,
|
|
1497 TrueCount);
|
|
1498 }
|
|
1499
|
|
1500 // If we have "X || 0", simplify the code to use an uncond branch.
|
|
1501 // "X || 1" would have been constant folded to 1.
|
|
1502 if (ConstantFoldsToSimpleInteger(CondBOp->getRHS(), ConstantBool) &&
|
|
1503 !ConstantBool) {
|
|
1504 // br(X || 0) -> br(X).
|
|
1505 return EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, FalseBlock,
|
|
1506 TrueCount);
|
|
1507 }
|
|
1508
|
|
1509 // Emit the LHS as a conditional. If the LHS conditional is true, we
|
|
1510 // want to jump to the TrueBlock.
|
|
1511 llvm::BasicBlock *LHSFalse = createBasicBlock("lor.lhs.false");
|
|
1512 // We have the count for entry to the RHS and for the whole expression
|
|
1513 // being true, so we can divy up True count between the short circuit and
|
|
1514 // the RHS.
|
|
1515 uint64_t LHSCount =
|
|
1516 getCurrentProfileCount() - getProfileCount(CondBOp->getRHS());
|
|
1517 uint64_t RHSCount = TrueCount - LHSCount;
|
|
1518
|
|
1519 ConditionalEvaluation eval(*this);
|
|
1520 {
|
|
1521 ApplyDebugLocation DL(*this, Cond);
|
|
1522 EmitBranchOnBoolExpr(CondBOp->getLHS(), TrueBlock, LHSFalse, LHSCount);
|
|
1523 EmitBlock(LHSFalse);
|
|
1524 }
|
|
1525
|
|
1526 incrementProfileCounter(CondBOp);
|
|
1527 setCurrentProfileCount(getProfileCount(CondBOp->getRHS()));
|
|
1528
|
|
1529 // Any temporaries created here are conditional.
|
|
1530 eval.begin(*this);
|
|
1531 EmitBranchOnBoolExpr(CondBOp->getRHS(), TrueBlock, FalseBlock, RHSCount);
|
|
1532
|
|
1533 eval.end(*this);
|
|
1534
|
|
1535 return;
|
|
1536 }
|
|
1537 }
|
|
1538
|
|
1539 if (const UnaryOperator *CondUOp = dyn_cast<UnaryOperator>(Cond)) {
|
|
1540 // br(!x, t, f) -> br(x, f, t)
|
|
1541 if (CondUOp->getOpcode() == UO_LNot) {
|
|
1542 // Negate the count.
|
|
1543 uint64_t FalseCount = getCurrentProfileCount() - TrueCount;
|
|
1544 // Negate the condition and swap the destination blocks.
|
|
1545 return EmitBranchOnBoolExpr(CondUOp->getSubExpr(), FalseBlock, TrueBlock,
|
|
1546 FalseCount);
|
|
1547 }
|
|
1548 }
|
|
1549
|
|
1550 if (const ConditionalOperator *CondOp = dyn_cast<ConditionalOperator>(Cond)) {
|
|
1551 // br(c ? x : y, t, f) -> br(c, br(x, t, f), br(y, t, f))
|
|
1552 llvm::BasicBlock *LHSBlock = createBasicBlock("cond.true");
|
|
1553 llvm::BasicBlock *RHSBlock = createBasicBlock("cond.false");
|
|
1554
|
|
1555 ConditionalEvaluation cond(*this);
|
|
1556 EmitBranchOnBoolExpr(CondOp->getCond(), LHSBlock, RHSBlock,
|
|
1557 getProfileCount(CondOp));
|
|
1558
|
|
1559 // When computing PGO branch weights, we only know the overall count for
|
|
1560 // the true block. This code is essentially doing tail duplication of the
|
|
1561 // naive code-gen, introducing new edges for which counts are not
|
|
1562 // available. Divide the counts proportionally between the LHS and RHS of
|
|
1563 // the conditional operator.
|
|
1564 uint64_t LHSScaledTrueCount = 0;
|
|
1565 if (TrueCount) {
|
|
1566 double LHSRatio =
|
|
1567 getProfileCount(CondOp) / (double)getCurrentProfileCount();
|
|
1568 LHSScaledTrueCount = TrueCount * LHSRatio;
|
|
1569 }
|
|
1570
|
|
1571 cond.begin(*this);
|
|
1572 EmitBlock(LHSBlock);
|
|
1573 incrementProfileCounter(CondOp);
|
|
1574 {
|
|
1575 ApplyDebugLocation DL(*this, Cond);
|
|
1576 EmitBranchOnBoolExpr(CondOp->getLHS(), TrueBlock, FalseBlock,
|
|
1577 LHSScaledTrueCount);
|
|
1578 }
|
|
1579 cond.end(*this);
|
|
1580
|
|
1581 cond.begin(*this);
|
|
1582 EmitBlock(RHSBlock);
|
|
1583 EmitBranchOnBoolExpr(CondOp->getRHS(), TrueBlock, FalseBlock,
|
|
1584 TrueCount - LHSScaledTrueCount);
|
|
1585 cond.end(*this);
|
|
1586
|
|
1587 return;
|
|
1588 }
|
|
1589
|
|
1590 if (const CXXThrowExpr *Throw = dyn_cast<CXXThrowExpr>(Cond)) {
|
|
1591 // Conditional operator handling can give us a throw expression as a
|
|
1592 // condition for a case like:
|
|
1593 // br(c ? throw x : y, t, f) -> br(c, br(throw x, t, f), br(y, t, f)
|
|
1594 // Fold this to:
|
|
1595 // br(c, throw x, br(y, t, f))
|
|
1596 EmitCXXThrowExpr(Throw, /*KeepInsertionPoint*/false);
|
|
1597 return;
|
|
1598 }
|
|
1599
|
|
1600 // If the branch has a condition wrapped by __builtin_unpredictable,
|
|
1601 // create metadata that specifies that the branch is unpredictable.
|
|
1602 // Don't bother if not optimizing because that metadata would not be used.
|
|
1603 llvm::MDNode *Unpredictable = nullptr;
|
|
1604 auto *Call = dyn_cast<CallExpr>(Cond->IgnoreImpCasts());
|
|
1605 if (Call && CGM.getCodeGenOpts().OptimizationLevel != 0) {
|
|
1606 auto *FD = dyn_cast_or_null<FunctionDecl>(Call->getCalleeDecl());
|
|
1607 if (FD && FD->getBuiltinID() == Builtin::BI__builtin_unpredictable) {
|
|
1608 llvm::MDBuilder MDHelper(getLLVMContext());
|
|
1609 Unpredictable = MDHelper.createUnpredictable();
|
|
1610 }
|
|
1611 }
|
|
1612
|
|
1613 // Create branch weights based on the number of times we get here and the
|
|
1614 // number of times the condition should be true.
|
|
1615 uint64_t CurrentCount = std::max(getCurrentProfileCount(), TrueCount);
|
|
1616 llvm::MDNode *Weights =
|
|
1617 createProfileWeights(TrueCount, CurrentCount - TrueCount);
|
|
1618
|
|
1619 // Emit the code with the fully general case.
|
|
1620 llvm::Value *CondV;
|
|
1621 {
|
|
1622 ApplyDebugLocation DL(*this, Cond);
|
|
1623 CondV = EvaluateExprAsBool(Cond);
|
|
1624 }
|
|
1625 Builder.CreateCondBr(CondV, TrueBlock, FalseBlock, Weights, Unpredictable);
|
|
1626 }
|
|
1627
|
|
1628 /// ErrorUnsupported - Print out an error that codegen doesn't support the
|
|
1629 /// specified stmt yet.
|
|
1630 void CodeGenFunction::ErrorUnsupported(const Stmt *S, const char *Type) {
|
|
1631 CGM.ErrorUnsupported(S, Type);
|
|
1632 }
|
|
1633
|
|
1634 /// emitNonZeroVLAInit - Emit the "zero" initialization of a
|
|
1635 /// variable-length array whose elements have a non-zero bit-pattern.
|
|
1636 ///
|
|
1637 /// \param baseType the inner-most element type of the array
|
|
1638 /// \param src - a char* pointing to the bit-pattern for a single
|
|
1639 /// base element of the array
|
|
1640 /// \param sizeInChars - the total size of the VLA, in chars
|
|
1641 static void emitNonZeroVLAInit(CodeGenFunction &CGF, QualType baseType,
|
|
1642 Address dest, Address src,
|
|
1643 llvm::Value *sizeInChars) {
|
|
1644 CGBuilderTy &Builder = CGF.Builder;
|
|
1645
|
|
1646 CharUnits baseSize = CGF.getContext().getTypeSizeInChars(baseType);
|
|
1647 llvm::Value *baseSizeInChars
|
|
1648 = llvm::ConstantInt::get(CGF.IntPtrTy, baseSize.getQuantity());
|
|
1649
|
|
1650 Address begin =
|
|
1651 Builder.CreateElementBitCast(dest, CGF.Int8Ty, "vla.begin");
|
|
1652 llvm::Value *end =
|
|
1653 Builder.CreateInBoundsGEP(begin.getPointer(), sizeInChars, "vla.end");
|
|
1654
|
|
1655 llvm::BasicBlock *originBB = CGF.Builder.GetInsertBlock();
|
|
1656 llvm::BasicBlock *loopBB = CGF.createBasicBlock("vla-init.loop");
|
|
1657 llvm::BasicBlock *contBB = CGF.createBasicBlock("vla-init.cont");
|
|
1658
|
|
1659 // Make a loop over the VLA. C99 guarantees that the VLA element
|
|
1660 // count must be nonzero.
|
|
1661 CGF.EmitBlock(loopBB);
|
|
1662
|
|
1663 llvm::PHINode *cur = Builder.CreatePHI(begin.getType(), 2, "vla.cur");
|
|
1664 cur->addIncoming(begin.getPointer(), originBB);
|
|
1665
|
|
1666 CharUnits curAlign =
|
|
1667 dest.getAlignment().alignmentOfArrayElement(baseSize);
|
|
1668
|
|
1669 // memcpy the individual element bit-pattern.
|
|
1670 Builder.CreateMemCpy(Address(cur, curAlign), src, baseSizeInChars,
|
|
1671 /*volatile*/ false);
|
|
1672
|
|
1673 // Go to the next element.
|
|
1674 llvm::Value *next =
|
|
1675 Builder.CreateInBoundsGEP(CGF.Int8Ty, cur, baseSizeInChars, "vla.next");
|
|
1676
|
|
1677 // Leave if that's the end of the VLA.
|
|
1678 llvm::Value *done = Builder.CreateICmpEQ(next, end, "vla-init.isdone");
|
|
1679 Builder.CreateCondBr(done, contBB, loopBB);
|
|
1680 cur->addIncoming(next, loopBB);
|
|
1681
|
|
1682 CGF.EmitBlock(contBB);
|
|
1683 }
|
|
1684
|
|
1685 void
|
|
1686 CodeGenFunction::EmitNullInitialization(Address DestPtr, QualType Ty) {
|
|
1687 // Ignore empty classes in C++.
|
|
1688 if (getLangOpts().CPlusPlus) {
|
|
1689 if (const RecordType *RT = Ty->getAs<RecordType>()) {
|
|
1690 if (cast<CXXRecordDecl>(RT->getDecl())->isEmpty())
|
|
1691 return;
|
|
1692 }
|
|
1693 }
|
|
1694
|
|
1695 // Cast the dest ptr to the appropriate i8 pointer type.
|
|
1696 if (DestPtr.getElementType() != Int8Ty)
|
|
1697 DestPtr = Builder.CreateElementBitCast(DestPtr, Int8Ty);
|
|
1698
|
|
1699 // Get size and alignment info for this aggregate.
|
|
1700 CharUnits size = getContext().getTypeSizeInChars(Ty);
|
|
1701
|
|
1702 llvm::Value *SizeVal;
|
|
1703 const VariableArrayType *vla;
|
|
1704
|
|
1705 // Don't bother emitting a zero-byte memset.
|
|
1706 if (size.isZero()) {
|
|
1707 // But note that getTypeInfo returns 0 for a VLA.
|
|
1708 if (const VariableArrayType *vlaType =
|
|
1709 dyn_cast_or_null<VariableArrayType>(
|
|
1710 getContext().getAsArrayType(Ty))) {
|
|
1711 auto VlaSize = getVLASize(vlaType);
|
|
1712 SizeVal = VlaSize.NumElts;
|
|
1713 CharUnits eltSize = getContext().getTypeSizeInChars(VlaSize.Type);
|
|
1714 if (!eltSize.isOne())
|
|
1715 SizeVal = Builder.CreateNUWMul(SizeVal, CGM.getSize(eltSize));
|
|
1716 vla = vlaType;
|
|
1717 } else {
|
|
1718 return;
|
|
1719 }
|
|
1720 } else {
|
|
1721 SizeVal = CGM.getSize(size);
|
|
1722 vla = nullptr;
|
|
1723 }
|
|
1724
|
|
1725 // If the type contains a pointer to data member we can't memset it to zero.
|
|
1726 // Instead, create a null constant and copy it to the destination.
|
|
1727 // TODO: there are other patterns besides zero that we can usefully memset,
|
|
1728 // like -1, which happens to be the pattern used by member-pointers.
|
|
1729 if (!CGM.getTypes().isZeroInitializable(Ty)) {
|
|
1730 // For a VLA, emit a single element, then splat that over the VLA.
|
|
1731 if (vla) Ty = getContext().getBaseElementType(vla);
|
|
1732
|
|
1733 llvm::Constant *NullConstant = CGM.EmitNullConstant(Ty);
|
|
1734
|
|
1735 llvm::GlobalVariable *NullVariable =
|
|
1736 new llvm::GlobalVariable(CGM.getModule(), NullConstant->getType(),
|
|
1737 /*isConstant=*/true,
|
|
1738 llvm::GlobalVariable::PrivateLinkage,
|
|
1739 NullConstant, Twine());
|
|
1740 CharUnits NullAlign = DestPtr.getAlignment();
|
|
1741 NullVariable->setAlignment(NullAlign.getAsAlign());
|
|
1742 Address SrcPtr(Builder.CreateBitCast(NullVariable, Builder.getInt8PtrTy()),
|
|
1743 NullAlign);
|
|
1744
|
|
1745 if (vla) return emitNonZeroVLAInit(*this, Ty, DestPtr, SrcPtr, SizeVal);
|
|
1746
|
|
1747 // Get and call the appropriate llvm.memcpy overload.
|
|
1748 Builder.CreateMemCpy(DestPtr, SrcPtr, SizeVal, false);
|
|
1749 return;
|
|
1750 }
|
|
1751
|
|
1752 // Otherwise, just memset the whole thing to zero. This is legal
|
|
1753 // because in LLVM, all default initializers (other than the ones we just
|
|
1754 // handled above) are guaranteed to have a bit pattern of all zeros.
|
|
1755 Builder.CreateMemSet(DestPtr, Builder.getInt8(0), SizeVal, false);
|
|
1756 }
|
|
1757
|
|
1758 llvm::BlockAddress *CodeGenFunction::GetAddrOfLabel(const LabelDecl *L) {
|
|
1759 // Make sure that there is a block for the indirect goto.
|
|
1760 if (!IndirectBranch)
|
|
1761 GetIndirectGotoBlock();
|
|
1762
|
|
1763 llvm::BasicBlock *BB = getJumpDestForLabel(L).getBlock();
|
|
1764
|
|
1765 // Make sure the indirect branch includes all of the address-taken blocks.
|
|
1766 IndirectBranch->addDestination(BB);
|
|
1767 return llvm::BlockAddress::get(CurFn, BB);
|
|
1768 }
|
|
1769
|
|
1770 llvm::BasicBlock *CodeGenFunction::GetIndirectGotoBlock() {
|
|
1771 // If we already made the indirect branch for indirect goto, return its block.
|
|
1772 if (IndirectBranch) return IndirectBranch->getParent();
|
|
1773
|
|
1774 CGBuilderTy TmpBuilder(*this, createBasicBlock("indirectgoto"));
|
|
1775
|
|
1776 // Create the PHI node that indirect gotos will add entries to.
|
|
1777 llvm::Value *DestVal = TmpBuilder.CreatePHI(Int8PtrTy, 0,
|
|
1778 "indirect.goto.dest");
|
|
1779
|
|
1780 // Create the indirect branch instruction.
|
|
1781 IndirectBranch = TmpBuilder.CreateIndirectBr(DestVal);
|
|
1782 return IndirectBranch->getParent();
|
|
1783 }
|
|
1784
|
|
1785 /// Computes the length of an array in elements, as well as the base
|
|
1786 /// element type and a properly-typed first element pointer.
|
|
1787 llvm::Value *CodeGenFunction::emitArrayLength(const ArrayType *origArrayType,
|
|
1788 QualType &baseType,
|
|
1789 Address &addr) {
|
|
1790 const ArrayType *arrayType = origArrayType;
|
|
1791
|
|
1792 // If it's a VLA, we have to load the stored size. Note that
|
|
1793 // this is the size of the VLA in bytes, not its size in elements.
|
|
1794 llvm::Value *numVLAElements = nullptr;
|
|
1795 if (isa<VariableArrayType>(arrayType)) {
|
|
1796 numVLAElements = getVLASize(cast<VariableArrayType>(arrayType)).NumElts;
|
|
1797
|
|
1798 // Walk into all VLAs. This doesn't require changes to addr,
|
|
1799 // which has type T* where T is the first non-VLA element type.
|
|
1800 do {
|
|
1801 QualType elementType = arrayType->getElementType();
|
|
1802 arrayType = getContext().getAsArrayType(elementType);
|
|
1803
|
|
1804 // If we only have VLA components, 'addr' requires no adjustment.
|
|
1805 if (!arrayType) {
|
|
1806 baseType = elementType;
|
|
1807 return numVLAElements;
|
|
1808 }
|
|
1809 } while (isa<VariableArrayType>(arrayType));
|
|
1810
|
|
1811 // We get out here only if we find a constant array type
|
|
1812 // inside the VLA.
|
|
1813 }
|
|
1814
|
|
1815 // We have some number of constant-length arrays, so addr should
|
|
1816 // have LLVM type [M x [N x [...]]]*. Build a GEP that walks
|
|
1817 // down to the first element of addr.
|
|
1818 SmallVector<llvm::Value*, 8> gepIndices;
|
|
1819
|
|
1820 // GEP down to the array type.
|
|
1821 llvm::ConstantInt *zero = Builder.getInt32(0);
|
|
1822 gepIndices.push_back(zero);
|
|
1823
|
|
1824 uint64_t countFromCLAs = 1;
|
|
1825 QualType eltType;
|
|
1826
|
|
1827 llvm::ArrayType *llvmArrayType =
|
|
1828 dyn_cast<llvm::ArrayType>(addr.getElementType());
|
|
1829 while (llvmArrayType) {
|
|
1830 assert(isa<ConstantArrayType>(arrayType));
|
|
1831 assert(cast<ConstantArrayType>(arrayType)->getSize().getZExtValue()
|
|
1832 == llvmArrayType->getNumElements());
|
|
1833
|
|
1834 gepIndices.push_back(zero);
|
|
1835 countFromCLAs *= llvmArrayType->getNumElements();
|
|
1836 eltType = arrayType->getElementType();
|
|
1837
|
|
1838 llvmArrayType =
|
|
1839 dyn_cast<llvm::ArrayType>(llvmArrayType->getElementType());
|
|
1840 arrayType = getContext().getAsArrayType(arrayType->getElementType());
|
|
1841 assert((!llvmArrayType || arrayType) &&
|
|
1842 "LLVM and Clang types are out-of-synch");
|
|
1843 }
|
|
1844
|
|
1845 if (arrayType) {
|
|
1846 // From this point onwards, the Clang array type has been emitted
|
|
1847 // as some other type (probably a packed struct). Compute the array
|
|
1848 // size, and just emit the 'begin' expression as a bitcast.
|
|
1849 while (arrayType) {
|
|
1850 countFromCLAs *=
|
|
1851 cast<ConstantArrayType>(arrayType)->getSize().getZExtValue();
|
|
1852 eltType = arrayType->getElementType();
|
|
1853 arrayType = getContext().getAsArrayType(eltType);
|
|
1854 }
|
|
1855
|
|
1856 llvm::Type *baseType = ConvertType(eltType);
|
|
1857 addr = Builder.CreateElementBitCast(addr, baseType, "array.begin");
|
|
1858 } else {
|
|
1859 // Create the actual GEP.
|
|
1860 addr = Address(Builder.CreateInBoundsGEP(addr.getPointer(),
|
|
1861 gepIndices, "array.begin"),
|
|
1862 addr.getAlignment());
|
|
1863 }
|
|
1864
|
|
1865 baseType = eltType;
|
|
1866
|
|
1867 llvm::Value *numElements
|
|
1868 = llvm::ConstantInt::get(SizeTy, countFromCLAs);
|
|
1869
|
|
1870 // If we had any VLA dimensions, factor them in.
|
|
1871 if (numVLAElements)
|
|
1872 numElements = Builder.CreateNUWMul(numVLAElements, numElements);
|
|
1873
|
|
1874 return numElements;
|
|
1875 }
|
|
1876
|
|
1877 CodeGenFunction::VlaSizePair CodeGenFunction::getVLASize(QualType type) {
|
|
1878 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
|
|
1879 assert(vla && "type was not a variable array type!");
|
|
1880 return getVLASize(vla);
|
|
1881 }
|
|
1882
|
|
1883 CodeGenFunction::VlaSizePair
|
|
1884 CodeGenFunction::getVLASize(const VariableArrayType *type) {
|
|
1885 // The number of elements so far; always size_t.
|
|
1886 llvm::Value *numElements = nullptr;
|
|
1887
|
|
1888 QualType elementType;
|
|
1889 do {
|
|
1890 elementType = type->getElementType();
|
|
1891 llvm::Value *vlaSize = VLASizeMap[type->getSizeExpr()];
|
|
1892 assert(vlaSize && "no size for VLA!");
|
|
1893 assert(vlaSize->getType() == SizeTy);
|
|
1894
|
|
1895 if (!numElements) {
|
|
1896 numElements = vlaSize;
|
|
1897 } else {
|
|
1898 // It's undefined behavior if this wraps around, so mark it that way.
|
|
1899 // FIXME: Teach -fsanitize=undefined to trap this.
|
|
1900 numElements = Builder.CreateNUWMul(numElements, vlaSize);
|
|
1901 }
|
|
1902 } while ((type = getContext().getAsVariableArrayType(elementType)));
|
|
1903
|
|
1904 return { numElements, elementType };
|
|
1905 }
|
|
1906
|
|
1907 CodeGenFunction::VlaSizePair
|
|
1908 CodeGenFunction::getVLAElements1D(QualType type) {
|
|
1909 const VariableArrayType *vla = getContext().getAsVariableArrayType(type);
|
|
1910 assert(vla && "type was not a variable array type!");
|
|
1911 return getVLAElements1D(vla);
|
|
1912 }
|
|
1913
|
|
1914 CodeGenFunction::VlaSizePair
|
|
1915 CodeGenFunction::getVLAElements1D(const VariableArrayType *Vla) {
|
|
1916 llvm::Value *VlaSize = VLASizeMap[Vla->getSizeExpr()];
|
|
1917 assert(VlaSize && "no size for VLA!");
|
|
1918 assert(VlaSize->getType() == SizeTy);
|
|
1919 return { VlaSize, Vla->getElementType() };
|
|
1920 }
|
|
1921
|
|
1922 void CodeGenFunction::EmitVariablyModifiedType(QualType type) {
|
|
1923 assert(type->isVariablyModifiedType() &&
|
|
1924 "Must pass variably modified type to EmitVLASizes!");
|
|
1925
|
|
1926 EnsureInsertPoint();
|
|
1927
|
|
1928 // We're going to walk down into the type and look for VLA
|
|
1929 // expressions.
|
|
1930 do {
|
|
1931 assert(type->isVariablyModifiedType());
|
|
1932
|
|
1933 const Type *ty = type.getTypePtr();
|
|
1934 switch (ty->getTypeClass()) {
|
|
1935
|
|
1936 #define TYPE(Class, Base)
|
|
1937 #define ABSTRACT_TYPE(Class, Base)
|
|
1938 #define NON_CANONICAL_TYPE(Class, Base)
|
|
1939 #define DEPENDENT_TYPE(Class, Base) case Type::Class:
|
|
1940 #define NON_CANONICAL_UNLESS_DEPENDENT_TYPE(Class, Base)
|
|
1941 #include "clang/AST/TypeNodes.inc"
|
|
1942 llvm_unreachable("unexpected dependent type!");
|
|
1943
|
|
1944 // These types are never variably-modified.
|
|
1945 case Type::Builtin:
|
|
1946 case Type::Complex:
|
|
1947 case Type::Vector:
|
|
1948 case Type::ExtVector:
|
173
|
1949 case Type::ConstantMatrix:
|
150
|
1950 case Type::Record:
|
|
1951 case Type::Enum:
|
|
1952 case Type::Elaborated:
|
|
1953 case Type::TemplateSpecialization:
|
|
1954 case Type::ObjCTypeParam:
|
|
1955 case Type::ObjCObject:
|
|
1956 case Type::ObjCInterface:
|
|
1957 case Type::ObjCObjectPointer:
|
173
|
1958 case Type::ExtInt:
|
150
|
1959 llvm_unreachable("type class is never variably-modified!");
|
|
1960
|
|
1961 case Type::Adjusted:
|
|
1962 type = cast<AdjustedType>(ty)->getAdjustedType();
|
|
1963 break;
|
|
1964
|
|
1965 case Type::Decayed:
|
|
1966 type = cast<DecayedType>(ty)->getPointeeType();
|
|
1967 break;
|
|
1968
|
|
1969 case Type::Pointer:
|
|
1970 type = cast<PointerType>(ty)->getPointeeType();
|
|
1971 break;
|
|
1972
|
|
1973 case Type::BlockPointer:
|
|
1974 type = cast<BlockPointerType>(ty)->getPointeeType();
|
|
1975 break;
|
|
1976
|
|
1977 case Type::LValueReference:
|
|
1978 case Type::RValueReference:
|
|
1979 type = cast<ReferenceType>(ty)->getPointeeType();
|
|
1980 break;
|
|
1981
|
|
1982 case Type::MemberPointer:
|
|
1983 type = cast<MemberPointerType>(ty)->getPointeeType();
|
|
1984 break;
|
|
1985
|
|
1986 case Type::ConstantArray:
|
|
1987 case Type::IncompleteArray:
|
|
1988 // Losing element qualification here is fine.
|
|
1989 type = cast<ArrayType>(ty)->getElementType();
|
|
1990 break;
|
|
1991
|
|
1992 case Type::VariableArray: {
|
|
1993 // Losing element qualification here is fine.
|
|
1994 const VariableArrayType *vat = cast<VariableArrayType>(ty);
|
|
1995
|
|
1996 // Unknown size indication requires no size computation.
|
|
1997 // Otherwise, evaluate and record it.
|
|
1998 if (const Expr *size = vat->getSizeExpr()) {
|
|
1999 // It's possible that we might have emitted this already,
|
|
2000 // e.g. with a typedef and a pointer to it.
|
|
2001 llvm::Value *&entry = VLASizeMap[size];
|
|
2002 if (!entry) {
|
|
2003 llvm::Value *Size = EmitScalarExpr(size);
|
|
2004
|
|
2005 // C11 6.7.6.2p5:
|
|
2006 // If the size is an expression that is not an integer constant
|
|
2007 // expression [...] each time it is evaluated it shall have a value
|
|
2008 // greater than zero.
|
|
2009 if (SanOpts.has(SanitizerKind::VLABound) &&
|
|
2010 size->getType()->isSignedIntegerType()) {
|
|
2011 SanitizerScope SanScope(this);
|
|
2012 llvm::Value *Zero = llvm::Constant::getNullValue(Size->getType());
|
|
2013 llvm::Constant *StaticArgs[] = {
|
|
2014 EmitCheckSourceLocation(size->getBeginLoc()),
|
|
2015 EmitCheckTypeDescriptor(size->getType())};
|
|
2016 EmitCheck(std::make_pair(Builder.CreateICmpSGT(Size, Zero),
|
|
2017 SanitizerKind::VLABound),
|
|
2018 SanitizerHandler::VLABoundNotPositive, StaticArgs, Size);
|
|
2019 }
|
|
2020
|
|
2021 // Always zexting here would be wrong if it weren't
|
|
2022 // undefined behavior to have a negative bound.
|
|
2023 entry = Builder.CreateIntCast(Size, SizeTy, /*signed*/ false);
|
|
2024 }
|
|
2025 }
|
|
2026 type = vat->getElementType();
|
|
2027 break;
|
|
2028 }
|
|
2029
|
|
2030 case Type::FunctionProto:
|
|
2031 case Type::FunctionNoProto:
|
|
2032 type = cast<FunctionType>(ty)->getReturnType();
|
|
2033 break;
|
|
2034
|
|
2035 case Type::Paren:
|
|
2036 case Type::TypeOf:
|
|
2037 case Type::UnaryTransform:
|
|
2038 case Type::Attributed:
|
|
2039 case Type::SubstTemplateTypeParm:
|
|
2040 case Type::PackExpansion:
|
|
2041 case Type::MacroQualified:
|
|
2042 // Keep walking after single level desugaring.
|
|
2043 type = type.getSingleStepDesugaredType(getContext());
|
|
2044 break;
|
|
2045
|
|
2046 case Type::Typedef:
|
|
2047 case Type::Decltype:
|
|
2048 case Type::Auto:
|
|
2049 case Type::DeducedTemplateSpecialization:
|
|
2050 // Stop walking: nothing to do.
|
|
2051 return;
|
|
2052
|
|
2053 case Type::TypeOfExpr:
|
|
2054 // Stop walking: emit typeof expression.
|
|
2055 EmitIgnoredExpr(cast<TypeOfExprType>(ty)->getUnderlyingExpr());
|
|
2056 return;
|
|
2057
|
|
2058 case Type::Atomic:
|
|
2059 type = cast<AtomicType>(ty)->getValueType();
|
|
2060 break;
|
|
2061
|
|
2062 case Type::Pipe:
|
|
2063 type = cast<PipeType>(ty)->getElementType();
|
|
2064 break;
|
|
2065 }
|
|
2066 } while (type->isVariablyModifiedType());
|
|
2067 }
|
|
2068
|
|
2069 Address CodeGenFunction::EmitVAListRef(const Expr* E) {
|
|
2070 if (getContext().getBuiltinVaListType()->isArrayType())
|
|
2071 return EmitPointerWithAlignment(E);
|
|
2072 return EmitLValue(E).getAddress(*this);
|
|
2073 }
|
|
2074
|
|
2075 Address CodeGenFunction::EmitMSVAListRef(const Expr *E) {
|
|
2076 return EmitLValue(E).getAddress(*this);
|
|
2077 }
|
|
2078
|
|
2079 void CodeGenFunction::EmitDeclRefExprDbgValue(const DeclRefExpr *E,
|
|
2080 const APValue &Init) {
|
|
2081 assert(Init.hasValue() && "Invalid DeclRefExpr initializer!");
|
|
2082 if (CGDebugInfo *Dbg = getDebugInfo())
|
|
2083 if (CGM.getCodeGenOpts().hasReducedDebugInfo())
|
|
2084 Dbg->EmitGlobalVariable(E->getDecl(), Init);
|
|
2085 }
|
|
2086
|
|
2087 CodeGenFunction::PeepholeProtection
|
|
2088 CodeGenFunction::protectFromPeepholes(RValue rvalue) {
|
|
2089 // At the moment, the only aggressive peephole we do in IR gen
|
|
2090 // is trunc(zext) folding, but if we add more, we can easily
|
|
2091 // extend this protection.
|
|
2092
|
|
2093 if (!rvalue.isScalar()) return PeepholeProtection();
|
|
2094 llvm::Value *value = rvalue.getScalarVal();
|
|
2095 if (!isa<llvm::ZExtInst>(value)) return PeepholeProtection();
|
|
2096
|
|
2097 // Just make an extra bitcast.
|
|
2098 assert(HaveInsertPoint());
|
|
2099 llvm::Instruction *inst = new llvm::BitCastInst(value, value->getType(), "",
|
|
2100 Builder.GetInsertBlock());
|
|
2101
|
|
2102 PeepholeProtection protection;
|
|
2103 protection.Inst = inst;
|
|
2104 return protection;
|
|
2105 }
|
|
2106
|
|
2107 void CodeGenFunction::unprotectFromPeepholes(PeepholeProtection protection) {
|
|
2108 if (!protection.Inst) return;
|
|
2109
|
|
2110 // In theory, we could try to duplicate the peepholes now, but whatever.
|
|
2111 protection.Inst->eraseFromParent();
|
|
2112 }
|
|
2113
|
173
|
2114 void CodeGenFunction::emitAlignmentAssumption(llvm::Value *PtrValue,
|
150
|
2115 QualType Ty, SourceLocation Loc,
|
|
2116 SourceLocation AssumptionLoc,
|
|
2117 llvm::Value *Alignment,
|
|
2118 llvm::Value *OffsetValue) {
|
|
2119 llvm::Value *TheCheck;
|
|
2120 llvm::Instruction *Assumption = Builder.CreateAlignmentAssumption(
|
|
2121 CGM.getDataLayout(), PtrValue, Alignment, OffsetValue, &TheCheck);
|
|
2122 if (SanOpts.has(SanitizerKind::Alignment)) {
|
173
|
2123 emitAlignmentAssumptionCheck(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
|
150
|
2124 OffsetValue, TheCheck, Assumption);
|
|
2125 }
|
|
2126 }
|
|
2127
|
173
|
2128 void CodeGenFunction::emitAlignmentAssumption(llvm::Value *PtrValue,
|
150
|
2129 const Expr *E,
|
|
2130 SourceLocation AssumptionLoc,
|
|
2131 llvm::Value *Alignment,
|
|
2132 llvm::Value *OffsetValue) {
|
|
2133 if (auto *CE = dyn_cast<CastExpr>(E))
|
|
2134 E = CE->getSubExprAsWritten();
|
|
2135 QualType Ty = E->getType();
|
|
2136 SourceLocation Loc = E->getExprLoc();
|
|
2137
|
173
|
2138 emitAlignmentAssumption(PtrValue, Ty, Loc, AssumptionLoc, Alignment,
|
150
|
2139 OffsetValue);
|
|
2140 }
|
|
2141
|
|
2142 llvm::Value *CodeGenFunction::EmitAnnotationCall(llvm::Function *AnnotationFn,
|
|
2143 llvm::Value *AnnotatedVal,
|
|
2144 StringRef AnnotationStr,
|
|
2145 SourceLocation Location) {
|
|
2146 llvm::Value *Args[4] = {
|
|
2147 AnnotatedVal,
|
|
2148 Builder.CreateBitCast(CGM.EmitAnnotationString(AnnotationStr), Int8PtrTy),
|
|
2149 Builder.CreateBitCast(CGM.EmitAnnotationUnit(Location), Int8PtrTy),
|
|
2150 CGM.EmitAnnotationLineNo(Location)
|
|
2151 };
|
|
2152 return Builder.CreateCall(AnnotationFn, Args);
|
|
2153 }
|
|
2154
|
|
2155 void CodeGenFunction::EmitVarAnnotations(const VarDecl *D, llvm::Value *V) {
|
|
2156 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
|
|
2157 // FIXME We create a new bitcast for every annotation because that's what
|
|
2158 // llvm-gcc was doing.
|
|
2159 for (const auto *I : D->specific_attrs<AnnotateAttr>())
|
|
2160 EmitAnnotationCall(CGM.getIntrinsic(llvm::Intrinsic::var_annotation),
|
|
2161 Builder.CreateBitCast(V, CGM.Int8PtrTy, V->getName()),
|
|
2162 I->getAnnotation(), D->getLocation());
|
|
2163 }
|
|
2164
|
|
2165 Address CodeGenFunction::EmitFieldAnnotations(const FieldDecl *D,
|
|
2166 Address Addr) {
|
|
2167 assert(D->hasAttr<AnnotateAttr>() && "no annotate attribute");
|
|
2168 llvm::Value *V = Addr.getPointer();
|
|
2169 llvm::Type *VTy = V->getType();
|
|
2170 llvm::Function *F = CGM.getIntrinsic(llvm::Intrinsic::ptr_annotation,
|
|
2171 CGM.Int8PtrTy);
|
|
2172
|
|
2173 for (const auto *I : D->specific_attrs<AnnotateAttr>()) {
|
|
2174 // FIXME Always emit the cast inst so we can differentiate between
|
|
2175 // annotation on the first field of a struct and annotation on the struct
|
|
2176 // itself.
|
|
2177 if (VTy != CGM.Int8PtrTy)
|
|
2178 V = Builder.CreateBitCast(V, CGM.Int8PtrTy);
|
|
2179 V = EmitAnnotationCall(F, V, I->getAnnotation(), D->getLocation());
|
|
2180 V = Builder.CreateBitCast(V, VTy);
|
|
2181 }
|
|
2182
|
|
2183 return Address(V, Addr.getAlignment());
|
|
2184 }
|
|
2185
|
|
2186 CodeGenFunction::CGCapturedStmtInfo::~CGCapturedStmtInfo() { }
|
|
2187
|
|
2188 CodeGenFunction::SanitizerScope::SanitizerScope(CodeGenFunction *CGF)
|
|
2189 : CGF(CGF) {
|
|
2190 assert(!CGF->IsSanitizerScope);
|
|
2191 CGF->IsSanitizerScope = true;
|
|
2192 }
|
|
2193
|
|
2194 CodeGenFunction::SanitizerScope::~SanitizerScope() {
|
|
2195 CGF->IsSanitizerScope = false;
|
|
2196 }
|
|
2197
|
|
2198 void CodeGenFunction::InsertHelper(llvm::Instruction *I,
|
|
2199 const llvm::Twine &Name,
|
|
2200 llvm::BasicBlock *BB,
|
|
2201 llvm::BasicBlock::iterator InsertPt) const {
|
|
2202 LoopStack.InsertHelper(I);
|
|
2203 if (IsSanitizerScope)
|
|
2204 CGM.getSanitizerMetadata()->disableSanitizerForInstruction(I);
|
|
2205 }
|
|
2206
|
|
2207 void CGBuilderInserter::InsertHelper(
|
|
2208 llvm::Instruction *I, const llvm::Twine &Name, llvm::BasicBlock *BB,
|
|
2209 llvm::BasicBlock::iterator InsertPt) const {
|
|
2210 llvm::IRBuilderDefaultInserter::InsertHelper(I, Name, BB, InsertPt);
|
|
2211 if (CGF)
|
|
2212 CGF->InsertHelper(I, Name, BB, InsertPt);
|
|
2213 }
|
|
2214
|
|
2215 static bool hasRequiredFeatures(const SmallVectorImpl<StringRef> &ReqFeatures,
|
|
2216 CodeGenModule &CGM, const FunctionDecl *FD,
|
|
2217 std::string &FirstMissing) {
|
|
2218 // If there aren't any required features listed then go ahead and return.
|
|
2219 if (ReqFeatures.empty())
|
|
2220 return false;
|
|
2221
|
|
2222 // Now build up the set of caller features and verify that all the required
|
|
2223 // features are there.
|
|
2224 llvm::StringMap<bool> CallerFeatureMap;
|
|
2225 CGM.getContext().getFunctionFeatureMap(CallerFeatureMap, FD);
|
|
2226
|
|
2227 // If we have at least one of the features in the feature list return
|
|
2228 // true, otherwise return false.
|
|
2229 return std::all_of(
|
|
2230 ReqFeatures.begin(), ReqFeatures.end(), [&](StringRef Feature) {
|
|
2231 SmallVector<StringRef, 1> OrFeatures;
|
|
2232 Feature.split(OrFeatures, '|');
|
|
2233 return llvm::any_of(OrFeatures, [&](StringRef Feature) {
|
|
2234 if (!CallerFeatureMap.lookup(Feature)) {
|
|
2235 FirstMissing = Feature.str();
|
|
2236 return false;
|
|
2237 }
|
|
2238 return true;
|
|
2239 });
|
|
2240 });
|
|
2241 }
|
|
2242
|
|
2243 // Emits an error if we don't have a valid set of target features for the
|
|
2244 // called function.
|
|
2245 void CodeGenFunction::checkTargetFeatures(const CallExpr *E,
|
|
2246 const FunctionDecl *TargetDecl) {
|
|
2247 return checkTargetFeatures(E->getBeginLoc(), TargetDecl);
|
|
2248 }
|
|
2249
|
|
2250 // Emits an error if we don't have a valid set of target features for the
|
|
2251 // called function.
|
|
2252 void CodeGenFunction::checkTargetFeatures(SourceLocation Loc,
|
|
2253 const FunctionDecl *TargetDecl) {
|
|
2254 // Early exit if this is an indirect call.
|
|
2255 if (!TargetDecl)
|
|
2256 return;
|
|
2257
|
|
2258 // Get the current enclosing function if it exists. If it doesn't
|
|
2259 // we can't check the target features anyhow.
|
|
2260 const FunctionDecl *FD = dyn_cast_or_null<FunctionDecl>(CurCodeDecl);
|
|
2261 if (!FD)
|
|
2262 return;
|
|
2263
|
|
2264 // Grab the required features for the call. For a builtin this is listed in
|
|
2265 // the td file with the default cpu, for an always_inline function this is any
|
|
2266 // listed cpu and any listed features.
|
|
2267 unsigned BuiltinID = TargetDecl->getBuiltinID();
|
|
2268 std::string MissingFeature;
|
|
2269 if (BuiltinID) {
|
|
2270 SmallVector<StringRef, 1> ReqFeatures;
|
|
2271 const char *FeatureList =
|
|
2272 CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
|
|
2273 // Return if the builtin doesn't have any required features.
|
|
2274 if (!FeatureList || StringRef(FeatureList) == "")
|
|
2275 return;
|
|
2276 StringRef(FeatureList).split(ReqFeatures, ',');
|
|
2277 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
|
|
2278 CGM.getDiags().Report(Loc, diag::err_builtin_needs_feature)
|
|
2279 << TargetDecl->getDeclName()
|
|
2280 << CGM.getContext().BuiltinInfo.getRequiredFeatures(BuiltinID);
|
|
2281
|
|
2282 } else if (!TargetDecl->isMultiVersion() &&
|
|
2283 TargetDecl->hasAttr<TargetAttr>()) {
|
|
2284 // Get the required features for the callee.
|
|
2285
|
|
2286 const TargetAttr *TD = TargetDecl->getAttr<TargetAttr>();
|
|
2287 ParsedTargetAttr ParsedAttr =
|
|
2288 CGM.getContext().filterFunctionTargetAttrs(TD);
|
|
2289
|
|
2290 SmallVector<StringRef, 1> ReqFeatures;
|
|
2291 llvm::StringMap<bool> CalleeFeatureMap;
|
173
|
2292 CGM.getContext().getFunctionFeatureMap(CalleeFeatureMap, TargetDecl);
|
150
|
2293
|
|
2294 for (const auto &F : ParsedAttr.Features) {
|
|
2295 if (F[0] == '+' && CalleeFeatureMap.lookup(F.substr(1)))
|
|
2296 ReqFeatures.push_back(StringRef(F).substr(1));
|
|
2297 }
|
|
2298
|
|
2299 for (const auto &F : CalleeFeatureMap) {
|
|
2300 // Only positive features are "required".
|
|
2301 if (F.getValue())
|
|
2302 ReqFeatures.push_back(F.getKey());
|
|
2303 }
|
|
2304 if (!hasRequiredFeatures(ReqFeatures, CGM, FD, MissingFeature))
|
|
2305 CGM.getDiags().Report(Loc, diag::err_function_needs_feature)
|
|
2306 << FD->getDeclName() << TargetDecl->getDeclName() << MissingFeature;
|
|
2307 }
|
|
2308 }
|
|
2309
|
|
2310 void CodeGenFunction::EmitSanitizerStatReport(llvm::SanitizerStatKind SSK) {
|
|
2311 if (!CGM.getCodeGenOpts().SanitizeStats)
|
|
2312 return;
|
|
2313
|
|
2314 llvm::IRBuilder<> IRB(Builder.GetInsertBlock(), Builder.GetInsertPoint());
|
|
2315 IRB.SetCurrentDebugLocation(Builder.getCurrentDebugLocation());
|
|
2316 CGM.getSanStats().create(IRB, SSK);
|
|
2317 }
|
|
2318
|
|
2319 llvm::Value *
|
|
2320 CodeGenFunction::FormResolverCondition(const MultiVersionResolverOption &RO) {
|
|
2321 llvm::Value *Condition = nullptr;
|
|
2322
|
|
2323 if (!RO.Conditions.Architecture.empty())
|
|
2324 Condition = EmitX86CpuIs(RO.Conditions.Architecture);
|
|
2325
|
|
2326 if (!RO.Conditions.Features.empty()) {
|
|
2327 llvm::Value *FeatureCond = EmitX86CpuSupports(RO.Conditions.Features);
|
|
2328 Condition =
|
|
2329 Condition ? Builder.CreateAnd(Condition, FeatureCond) : FeatureCond;
|
|
2330 }
|
|
2331 return Condition;
|
|
2332 }
|
|
2333
|
|
2334 static void CreateMultiVersionResolverReturn(CodeGenModule &CGM,
|
|
2335 llvm::Function *Resolver,
|
|
2336 CGBuilderTy &Builder,
|
|
2337 llvm::Function *FuncToReturn,
|
|
2338 bool SupportsIFunc) {
|
|
2339 if (SupportsIFunc) {
|
|
2340 Builder.CreateRet(FuncToReturn);
|
|
2341 return;
|
|
2342 }
|
|
2343
|
|
2344 llvm::SmallVector<llvm::Value *, 10> Args;
|
|
2345 llvm::for_each(Resolver->args(),
|
|
2346 [&](llvm::Argument &Arg) { Args.push_back(&Arg); });
|
|
2347
|
|
2348 llvm::CallInst *Result = Builder.CreateCall(FuncToReturn, Args);
|
|
2349 Result->setTailCallKind(llvm::CallInst::TCK_MustTail);
|
|
2350
|
|
2351 if (Resolver->getReturnType()->isVoidTy())
|
|
2352 Builder.CreateRetVoid();
|
|
2353 else
|
|
2354 Builder.CreateRet(Result);
|
|
2355 }
|
|
2356
|
|
2357 void CodeGenFunction::EmitMultiVersionResolver(
|
|
2358 llvm::Function *Resolver, ArrayRef<MultiVersionResolverOption> Options) {
|
|
2359 assert(getContext().getTargetInfo().getTriple().isX86() &&
|
|
2360 "Only implemented for x86 targets");
|
|
2361
|
|
2362 bool SupportsIFunc = getContext().getTargetInfo().supportsIFunc();
|
|
2363
|
|
2364 // Main function's basic block.
|
|
2365 llvm::BasicBlock *CurBlock = createBasicBlock("resolver_entry", Resolver);
|
|
2366 Builder.SetInsertPoint(CurBlock);
|
|
2367 EmitX86CpuInit();
|
|
2368
|
|
2369 for (const MultiVersionResolverOption &RO : Options) {
|
|
2370 Builder.SetInsertPoint(CurBlock);
|
|
2371 llvm::Value *Condition = FormResolverCondition(RO);
|
|
2372
|
|
2373 // The 'default' or 'generic' case.
|
|
2374 if (!Condition) {
|
|
2375 assert(&RO == Options.end() - 1 &&
|
|
2376 "Default or Generic case must be last");
|
|
2377 CreateMultiVersionResolverReturn(CGM, Resolver, Builder, RO.Function,
|
|
2378 SupportsIFunc);
|
|
2379 return;
|
|
2380 }
|
|
2381
|
|
2382 llvm::BasicBlock *RetBlock = createBasicBlock("resolver_return", Resolver);
|
|
2383 CGBuilderTy RetBuilder(*this, RetBlock);
|
|
2384 CreateMultiVersionResolverReturn(CGM, Resolver, RetBuilder, RO.Function,
|
|
2385 SupportsIFunc);
|
|
2386 CurBlock = createBasicBlock("resolver_else", Resolver);
|
|
2387 Builder.CreateCondBr(Condition, RetBlock, CurBlock);
|
|
2388 }
|
|
2389
|
|
2390 // If no generic/default, emit an unreachable.
|
|
2391 Builder.SetInsertPoint(CurBlock);
|
|
2392 llvm::CallInst *TrapCall = EmitTrapCall(llvm::Intrinsic::trap);
|
|
2393 TrapCall->setDoesNotReturn();
|
|
2394 TrapCall->setDoesNotThrow();
|
|
2395 Builder.CreateUnreachable();
|
|
2396 Builder.ClearInsertionPoint();
|
|
2397 }
|
|
2398
|
|
2399 // Loc - where the diagnostic will point, where in the source code this
|
|
2400 // alignment has failed.
|
|
2401 // SecondaryLoc - if present (will be present if sufficiently different from
|
|
2402 // Loc), the diagnostic will additionally point a "Note:" to this location.
|
|
2403 // It should be the location where the __attribute__((assume_aligned))
|
|
2404 // was written e.g.
|
173
|
2405 void CodeGenFunction::emitAlignmentAssumptionCheck(
|
150
|
2406 llvm::Value *Ptr, QualType Ty, SourceLocation Loc,
|
|
2407 SourceLocation SecondaryLoc, llvm::Value *Alignment,
|
|
2408 llvm::Value *OffsetValue, llvm::Value *TheCheck,
|
|
2409 llvm::Instruction *Assumption) {
|
|
2410 assert(Assumption && isa<llvm::CallInst>(Assumption) &&
|
173
|
2411 cast<llvm::CallInst>(Assumption)->getCalledOperand() ==
|
150
|
2412 llvm::Intrinsic::getDeclaration(
|
|
2413 Builder.GetInsertBlock()->getParent()->getParent(),
|
|
2414 llvm::Intrinsic::assume) &&
|
|
2415 "Assumption should be a call to llvm.assume().");
|
|
2416 assert(&(Builder.GetInsertBlock()->back()) == Assumption &&
|
|
2417 "Assumption should be the last instruction of the basic block, "
|
|
2418 "since the basic block is still being generated.");
|
|
2419
|
|
2420 if (!SanOpts.has(SanitizerKind::Alignment))
|
|
2421 return;
|
|
2422
|
|
2423 // Don't check pointers to volatile data. The behavior here is implementation-
|
|
2424 // defined.
|
|
2425 if (Ty->getPointeeType().isVolatileQualified())
|
|
2426 return;
|
|
2427
|
|
2428 // We need to temorairly remove the assumption so we can insert the
|
|
2429 // sanitizer check before it, else the check will be dropped by optimizations.
|
|
2430 Assumption->removeFromParent();
|
|
2431
|
|
2432 {
|
|
2433 SanitizerScope SanScope(this);
|
|
2434
|
|
2435 if (!OffsetValue)
|
|
2436 OffsetValue = Builder.getInt1(0); // no offset.
|
|
2437
|
|
2438 llvm::Constant *StaticData[] = {EmitCheckSourceLocation(Loc),
|
|
2439 EmitCheckSourceLocation(SecondaryLoc),
|
|
2440 EmitCheckTypeDescriptor(Ty)};
|
|
2441 llvm::Value *DynamicData[] = {EmitCheckValue(Ptr),
|
|
2442 EmitCheckValue(Alignment),
|
|
2443 EmitCheckValue(OffsetValue)};
|
|
2444 EmitCheck({std::make_pair(TheCheck, SanitizerKind::Alignment)},
|
|
2445 SanitizerHandler::AlignmentAssumption, StaticData, DynamicData);
|
|
2446 }
|
|
2447
|
|
2448 // We are now in the (new, empty) "cont" basic block.
|
|
2449 // Reintroduce the assumption.
|
|
2450 Builder.Insert(Assumption);
|
|
2451 // FIXME: Assumption still has it's original basic block as it's Parent.
|
|
2452 }
|
|
2453
|
|
2454 llvm::DebugLoc CodeGenFunction::SourceLocToDebugLoc(SourceLocation Location) {
|
|
2455 if (CGDebugInfo *DI = getDebugInfo())
|
|
2456 return DI->SourceLocToDebugLoc(Location);
|
|
2457
|
|
2458 return llvm::DebugLoc();
|
|
2459 }
|